Tracking: stabilize parallel compilation features

[majocha]

to eventually take them out of preview.

Currently the whole thing is enabled for LangVersion=preview: #17948

Prior work:
#17442
#17400

At the moment LangVersion preview does not guarantee full test coverage in the test suite.
The plan is to experimentally force enable each feature, to exercise it over all of the tests, uncover and fix issues.

Issues so far:

• Non deterministic edge case with parallel reference resolution #18977 (parallel resolution) - resolved in WIP: Parallel reference resolution with deterministic order #18985
• graph type checking failing in fsharpQA (E-SignatureAfterSource) - fixed by adding a reverse dependency .fsi -> .fs for such cases.
• graph checking of script compilation with #load directives (several test cases cover it) - fixed by processing the load closure part in sequence.
• sus Index was outside the bounds of the array and other errors in EVIL_PROVIDER_ReturnsTypeWithIncorrectNameFromAppl test . Hard to reproduce. Hopefully fixed by adding a missing lock.

What works:

• Optimization - no problems. Experiment: enable parallel optimization by default #18986

• IL gen - just some reordered methods, require baseline update. Experiment: enable parallel optimization by default #18986

[majocha]

All the fixes merged together in #18998.

[T-Gro]

We should keep switch(es) to opt-out of this if any not-yet-observed problem occurs.