[Incremental] Supports per-type-body fingerprints, on-by-default

[davidungar]

1. Changed granularity of (fine-grained) dependency graph traversal to support tracking changes to individual node fingerprints.
2. Changed semantics of transitive traversal to simplify meaning of a mark from "was a dependent of a cascading job" to "this node has been traversed".
3. Adds -enable-type-fingerprints disabled by default.
4. Changes driver handling of jobs when one job fails. Don't bother marking all unfinished jobs as cascading if they were previously thought to be. Just mark then as non-cascading-build required. On next run, if they are cascading the dependent jobs will get run.
5. When the option above is set, the parser will hash tokens within a nominal type body or extension body into a per-body hash instead of the whole-file interface hash.
6. Adjust tests accordingly.

-- A virtually-identical PR, 29482, has been examined by @CodaFi

[davidungar]

@swift-ci please smoke test os x platform

[davidungar]

@swift-ci please smoke test os x platform

[davidungar]

@swift-ci please clean smoke test os x platform

[davidungar]

@swift-ci please smoke test os x platform

[davidungar]

@swift-ci please smoke test

[davidungar]

@swift-ci please smoke test

[davezarzycki]

Why MD5? Why not SHA1 or SHA256 which can be hardware accelerated on some processors? (Yes, SHA1 is slightly bigger, but you can just truncate the result if that matters.)

[CodaFi]

I imagine it’s because LLVM already has the facilities for this. I’d prefer we use something like 128-bit siphash but it’s certainly a decision we can revisit.

[davidungar]

Yes, the use of MD5 follows what was done for the original, whole-file, interface hash.