[FEA] Provide cuOpt builds with CUOPT_INSTANTIATE_INT64 enabled

In order to make the GAMS<->cuOpt link (see also) work with very large problems with more than 2^31 non-zeroes (which GAMS supports with other solvers like, for example, CPLEX already), it would be helpful to have a prebuilt cuOpt package available from pip that has CUOPT_INSTANTIATE_INT64=1. I attempted to build it myself in a fork and that sort of worked and produced a wheel but I struggle to integrate this into the link.

Hence I would be interested in some instructions for specifically building cuOpt with int64 and using that in conjunction with the most recent nightly build wheels OR even better have int64-enabled packages directly available in the NVIDIA repositories. I understand making INT64 the default doesn't make sense as it would blow up the memory footprint (and hence slow down processing) even when INT32 is perfectly fine (for smaller problems).