Reduce cabalProject IFD deps using dummy GHCs

[hamishmack]

The cabalProject function works by running cabal v2-configure followed by plan-to-nix. However for v2-configure to correctly calculate the plan (including flag settings and buildable components) it needs to know what GHC we want to use.

This causes significant problems:

• For hydra to assemble a list of jobs from components.tests it must first have GHC that will be used. If a patch has been applied to the GHC to be used it must be rebuilt before the list of jobs can be assembled. If a lot of different GHCs are being tests that can be a lot of work all happening in the eval stage where little feedback is available.

• Once the jobs are running the compilation of the GHC needed (the eval stage already must have done it, but the outputs there are apparently not added to the cache) happens inside the IFD part of cabalProject. This causes a very large amount of work to be done in the IFD and our understanding is that this can cause problems on nix and/or hydra.

• When using cabalProject we cannot examine the properties of the project without building or downloading the GHC (less of an issue as we would normally need it soon anyway).

The solution here is to:

• Create a dummy ghc and ghc-pkg from the real ones by capturing the ghc and ghc-pkg outputs that cabal v2-configure requests.

• The dummy ghc and ghc-pkg is used instead of the real one in cabalProject when running cabal v2-configure.

• The captured outputs are materialized into the haskell.nix repo so that the real GHC is only needed when checkMaterialization is set.

[rvl]

This is quite impressive @hamishmack 😄