Existence of fine moduli space for curves and elliptic curves 
*

*For the moduli problem of a curve of genus $g$ with $n$ marked points, how large an $n$ is needed to ensure the existence of a fine moduli space? For this question, terminology is that of Mumford's GIT.

*For the following three moduli problems, how big an $N$ is required for existence of a fine moduli space? The terminology is from the exposes of Deligne-Rapoport and Katz-Mazur, or Shimura. The first is in French, the second is too big, and the third is using old language and never mentions the modern terminology of universal elliptic curve, etc.. Therefore it is not possible for me to dig up the information myself.
i) Elliptic curves equipped with a cyclic subgroup of order $N$ -- this moduli problem corresponds to the modular group $\Gamma_0(N)$.
ii) Elliptic curves equipped with a point of order $N$ -- this moduli problem corresponds to the modular group $\Gamma_1(N)$.
ii) Elliptic curves equipped with a symplectic pairing on $N$-torsion points -- this moduli problem corresponds to the modular group $\Gamma (N)$.
References other than the above, will be appreciated.
 A: Here is a thought on the first question.  What you need to know (at least to get an algebraic space; I'll let others be more careful than I if you want a scheme) is how large n must be to ensure that an automorphism of a smooth genus g curve X which fixes n points must be the identity.  Let G be the cyclic group generated by this automorphism:  then the map X -> X/G is totally ramified at your n fixed points.  So by Riemann-Hurwitz, g(X) [NO, 2g(X)-2, THANKS, BJORN) is at least -2|G| + n(|G|-1).  If G is nontrivial, in other words, g is at least n-4 [NO, 2g+2, THANKS, BJORN].  So I think g+5 [NO, 2g+3, THANKS, BJORN] marked points should be enough.  That this is necessary can be seen by taking g=2; on M_{2,6} you'll have a bunch of loci with an extra involution, parametrizing curves whose marked points are precisely the Weierstrass points.
[NO MORE LATE-NIGHT RIEMANN-HURWITZ:  THANKS TO BJORN FOR CORRECTING THE ERRORS]
A: The first is unrepresentable for arbitrary large $N$ (it depends on the residue class of 
$N$ mod 12), the second is representatble for $N \geq 4$ (if you are considering $Y_1(N)$)
or $N \geq 5$ (if you are considering $X_1(N)$, i.e. including the cusps), the third 
is representable for $N \geq 3$.  
The references you mentioned are the standard ones.  Probably Silverman discusses these in his books somewhere too (maybe the 2nd).  If you look in Gross's Duke paper on companion forms (A tameness criterion ... ) you will find a summary of the story for $X_1(N)$. In the
$\Gamma_0(N)$ case, Mazur has a careful discussion in the beginning of section 2 of his Eisenstein ideal paper.  Both Gross and Mazur refer back to Deligne--Rapoport
for proofs.
It is also just a matter of computing the torsion in each of the $\Gamma$'s (plus epsilon
more if you want to understand representability at the cusps), which is an exercise.  (Although you have to do a little work to see why this is the necessary computation.)
A: If you want to work over a base ring such as $\mathbf{Z}[1/n]$ rather than over $\mathbf{Q}$ or $\mathbf{C}$ then the relevant numerical condition is that the part of $N$ coprime to $n$ not be "too small" in the $\Gamma_1$ and full level cases.  For an extreme example, if $N$ is a $p$-power and you work over $\mathbf{Z}_{(p)}$ then you'll always have problems in characteristic $p$ at the supersingular points.
On the other hand, if you're willing to go beyond schemes and work with algebraic spaces or Deligne-Mumford or Artin stacks then these issues go away (at the expense of more technical background) in the sense that one has a reasonable "moduli space" over $\mathbf{Z}$ with nice regularity properties for all $N$ (even incorporating degenerations in the sense of generalized elliptic curves with level structure). It has better properties than a coarse moduli space (aside from perhaps not being a scheme...).
