It turns out that, as spin suggested in the comments, representing elements of $\mathrm{Co}_0$ as permutations of $196\,560$ elements (the first shell of the Leech lattice) is computationally feasible and that, with this representation, GAP is able to list the conjugacy classes, which are then easily converted back to matrices. I wrote the following code (expanding on the one given in the question) which computes the list of representative matrices — in less than half an hour on my PC — and dumps it out to a file:

```
#### Start gap with enough memory (e.g., "gap -o 8g" for 8GB) to run this.
## The MOG cells are labeled top to bottom and left to right:
## so the leftmost column is 1,2,3,4 and the rightmost is 21,22,23,24,
## whereas the top line is 1,5,9,13,17,21.
## Generators of the Mathieu group M_24:
rotate_blocks := (1,9,17)(2,10,18)(3,11,19)(4,12,20)(5,13,21)(6,14,22)(7,15,23)(8,16,24);
flip_blocks12 := (1,5)(2,6)(3,7)(4,8)(9,13)(10,14)(11,15)(12,16);
omega_cols := (2,3,4)(6,7,8)(10,11,12)(14,15,16)(18,19,20)(22,23,24);
myflip := (1,8)(2,4)(3,5)(6,7)(9,17)(10,18)(13,21)(14,22);
m24_group := Group([rotate_blocks, flip_blocks12, omega_cols, myflip]);
## The same permutations as matrices:
rotate_blocks_mat := PermutationMat(rotate_blocks, 24);
flip_blocks12_mat := PermutationMat(flip_blocks12, 24);
omega_cols_mat := PermutationMat(omega_cols, 24);
myflip_mat := PermutationMat(myflip, 24);
## Conway's xi generator first subtracts from each column half the sum
## of all its cells (this is eta), then takes the negative of one block:
conway_eta := List([1..24], i->List([1..24], function(j) if QuoInt(i-1,4)=QuoInt(j-1,4) then if i=j then return 1/2; else return -1/2; fi; else return 0; fi; end));;
conway_xi := conway_eta * List([1..24], i->List([1..24], function(j) if i=j then if i>=1 and i<=4 then return -1; else return 1; fi; else return 0; fi; end));
## Conway's group Co_0 as a group of matrices (too unwieldly for GAP):
co0_mat := Group([rotate_blocks_mat, flip_blocks12_mat, omega_cols_mat, myflip_mat, conway_xi]);
## The orbit of an element of the Leech lattice (this should have size 196560):
orb := Orbit(co0_mat, List([1..24], function(i) if i=1 or i=5 then return 4; else return 0; fi; end));;
Size(orb); ## Should return 196560
## The action homomorphism (makes elements of Co_0 into permutations):
phi := ActionHomomorphism(co0_mat, orb);
## The five generators, this time as permutations on orb (of size 196560):
rotate_blocks_l := Image(phi, rotate_blocks_mat);;
flip_blocks12_l := Image(phi, flip_blocks12_mat);;
omega_cols_l := Image(phi, omega_cols_mat);;
myflip_l := Image(phi, myflip_mat);;
conway_xi_l := Image(phi, conway_xi);;
## Now Co_0 as a permutation group:
co0 := Group([rotate_blocks_l, flip_blocks12_l, omega_cols_l, myflip_l, conway_xi_l]);
## Check its order and compute its conjugacy classes (this is fairly long):
Order(co0); ## Should return 8315553613086720000
lst_cl := ConjugacyClasses(co0);; ## Takes about 35* longer than previous command
lst := List(lst_cl, cl->Representative(cl));;
## Convenience function for cycle structure of permutations:
cycle_struct := function(p) local str, t, i; t := []; if NrMovedPoints(p)<Size(orb) then Append(t,[[1, Size(orb)-NrMovedPoints(p)]]); fi; str := CycleStructurePerm(p); for i in [1..Length(str)] do if IsBound(str[i]) then Append(t,[[i+1, str[i]]]); fi; od; return t; end;
## Morphisms from a free group to our two descriptions of Co_0:
frhom := EpimorphismFromFreeGroup(co0 : names := ["rot","fbl","omg","flp","cxi"]);;
frhom_mat := GroupHomomorphismByImages(Source(frhom), co0_mat, GeneratorsOfGroup(co0_mat));;
## Convert the computed conjugacy classes into representative matrices:
lst_mat := List(lst, x->Image(frhom_mat, PreImagesRepresentative(frhom, x)));;
## Now we try to identify the ATLAS labeling of these classes...
## Identify each class by its order and the traces of some powers:
keylist1 := List(lst_mat, m->[Order(m), Trace(m), Trace(m^2), Trace(m^3), Trace(m^4), Trace(m^6)]);;
## Compute centralizer orders (this is again a bit long):
lst_centralizers := List(lst, x->Order(Centralizer(co0, x)));;
## Get the ATLAS character table from the library:
tbl := CharacterTable("2.Co1");
# ConnectGroupAndCharacterTable(co0, tbl); ## <- This doesn't work, of course
## Identify the 24-dimensional character
tbl_stdchar := Filtered(Irr(tbl), l->l[1]=24)[1];
## Identify the classes by the same numbers, in the order given by tbl:
keylist2 := List([1..Length(lst_cl)], i->[OrdersClassRepresentatives(tbl)[i], tbl_stdchar[i], tbl_stdchar[PowerMap(tbl,2)[i]], tbl_stdchar[PowerMap(tbl,3)[i]], tbl_stdchar[PowerMap(tbl,4)[i]], tbl_stdchar[PowerMap(tbl,6)[i]]]);;
## Now recover the labeling, insofar as it can be identified:
labeling := [];; revorder := [];; for i in [1..Length(lst_cl)] do tmp := Filtered([1..Length(lst_cl)], j->keylist2[j]=keylist1[i]); if Length(tmp)=1 then labeling[i] := AtlasClassNames(tbl)[tmp[1]]; elif Length(tmp)=2 then labeling[i] := Concatenation(AtlasClassNames(tbl)[tmp[1]],"|",AtlasClassNames(tbl)[tmp[2]]); fi; if IsBound(revorder[tmp[1]]) then revorder[tmp[2]] := i; else revorder[tmp[1]] := i; fi; od;
## Finally, dump it all out il a file called "co0-representative-matrices.dat":
output := OutputTextFile("co0-representative-matrices.dat", false);
PrintTo(output, "lst_mat := [\n\n");
for j in [1..Length(lst_cl)] do i := revorder[j]; WriteAll(output, Concatenation("#### ", String(j), "\n## Label: ", labeling[i], "\n## Order: ", String(keylist1[i][1]), "\n## Centralizer: ", String(lst_centralizers[i]), "\n## Cycles: ", String(cycle_struct(lst[i])), "\n## Trace: ", String(keylist1[i][2]), "\n")); PrintTo(output, lst_mat[i], "\n,\n\n"); od;
PrintTo(output, "];\n\n");
PrintTo(output, "########\n\n");
PrintTo(output, "lst_labels := ", List(revorder, i->labeling[i]), ";\n\n");
PrintTo(output, "lst_orders := ", List(revorder, i->keylist1[i][1]), ";\n\n");
PrintTo(output, "lst_centralizers := ", List(revorder, i->lst_centralizers[i]), ";\n\n");
PrintTo(output, "lst_cycles := ", List(revorder, i->cycle_struct(lst[i])), ";\n\n");
PrintTo(output, "lst_traces := ", List(revorder, i->keylist1[i][2]), ";\n\n");
CloseStream(output);
```

I also placed the same code on GitHub in case this is more convenient. (And sorry if my identifier names are atrocious.)

The result of the computation can be downloaded here (very slightly hand-edited in comparison to the file output by the above code, so as to make an arbitrary choice of labeling for those classes that cannot be unambiguously identified w.r.t. the ATLAS). It is itself a valid GAP file.

base: for example $B = \{1,2,3,5,6134\}$ for a base. Here $B = \{b_1, \ldots, b_n\}$ is a base if $\cap_{i = 1}^n G_{b_i} = 1$. Then the action of any $g \in G$ is determined by $g(b)$ for $b \in B$, so by action on just five points in this case. Also, you can compute the order as $|G| = \prod_{i = 1}^n[G_{i-1}:G_{i}]$ where $G_0 = G$ and $G_i = \cap_{j = 1}^i G_{b_j}$ for $i > 0$. $\endgroup$