I can give a partial answer to the first question:

While the general problem of detecting
a Hamiltonian path or cycle on an
undirected grid graph is known to be
NP-complete, are there interesting
special cases where efficient
polynomial time algorithms exist for
enumerating all such paths/cycles?

If the grid graph is "solid," ie., has no holes, then there is a polynomial-time algorithm by Umans and Lenhart (paper here) that will find a Hamiltonian cycle, or reject the graph if no such cycle exists. The algorithm first finds a maximum matching, and then decomposes the graph into "static alternating strips," both of which can be performed efficiently. Production of the Hamiltonian cycle is achieved by changing the matching depending on how the static alternating strips are laid out.

While there may be exponentially many different H cycles, it is possible to enumerate them with polynomial delay (meaning only having to wait for a polynomial amount of time before outputting the next one) by changing the order in which one traverses the static alternating strips, and/or changing the underlying matching. (Caveat: the enumeration algorithm may need to be more careful than my handwaving, to ensure only polynomially-many duplicate cycles are outputted before a new one is. It seems, though, that one could simply build different cycles in parallel, and then prioritize the ones that deviate from one another.)

So hole-free grid graphs appear to be one such special case.