Integrate widh CompilerCaching.jl

[maleadt]

Some refactoring as well, so the diff is messier than expected, but it does show the intended API usage. Structured IR is saved as the "IR", Tile IR bitcode as "code", and CuFunctions as session-dependent "executables". We could actually store CUBINs in the "code" phase, but tileiras isn't always available (only on CUDA platforms), so to let more of the tests work I've put that in the final link phase where CUDA.jl is expected to be available.

One point of discussion is that code_structured etc now return a single entry as we rely on the caching infrastructure to do what would have happened at run time, whereas InteractiveUtils.code_... can return multiple matches. @vchuravy Thoughts?

[vchuravy]

One point of discussion is that code_structured etc now return a single entry as we rely on the caching infrastructure to do what would have happened at run time, whereas InteractiveUtils.code_... can return multiple matches.

The reason why they return a list is because the methods query can return multiple answers. It is sometimes useful if you want to look at multiple methods at once.