Optimize cached-build codepaths

[lihaoyi]

Running time ./mill __.compile on the Mill codebase, after everything is already compiled, takes about ~2.5 to run and do nothing. ./mill plan __.compile shows about 16847 tasks, working out to be able 0.14ms to evaluate each cached task.

Mill's codebase is unusual in having a large number of modules for its size, as it uses modules to define the various integration test cases. But this large number of modules is perhaps representative of large projects in general, and we should definitely be able to have better performance than this to support large projects.

[lihaoyi]

Some rough profiles showing the hotspots in JProfiler. Seems a lot of time spend in revalidateIfNeededOrThrow and in resolveDirectChildren (which uses Class#getMethods and NameTransformer#decode

Untitled.jps.zip

[lihaoyi]

Seems the time is relatively evenly split between resolve, instantiateModule, and evaluate

Optimizing resolve would likely involve moving the getMethods/decode calls into the Discover data structure, which would be a binary compatibility breakage and need to wait for 0.13.0