Various antigen improvements

[dhartglassMSFT]

Some fixes to some issues in the pipeline, and to make Antigen produce more useful testcases

Pipeline improvements

• Antigen used to reflect over its own runtime for intrinsic names. Caused bad testcase generation for arm64 where some of those experimental APIs got renamed. Also caused us to not test the latest intrinsics. Instead change to use MetadataLoadContext to load methods from the supplied Core_Root.
• Linux/MAC runs would never run to completion. This was because testcase's callstacks were too small, tests would stack overflow, and this would take down Antigen process itself. Fix by eunning testcases inside a dedicated thread with a larger stack, as Fuzzlyn currently also does.
• Detect debug/release diffs using a run-to-run stable hash algorithm instead of *GetHashCode() which is only stable per process. This wasn't "broken", because base+diff runs both executed ion the same process, but this should make these repros much easier to work with.
• Fix an issue with bucketing for runtime errors.

Testcase improvements

• Reduce probability of *.AllBitsSet() (NaN) on floating point vectors from 50% to 5%. Hopefully not having NaN everywhere will make for more interesting computations.
• Disallow Floating point->Integral reinterpret casts. This should avoid a common false positive where builds differ due to the bitwise representation of a NaN, which IS allowed to differ in debug and release builds.