[Hashing] Replace CityHash mixers with xxh3 (#194567)

Replace the CityHash-style mixer in hash_combine and (transitively)
hash_value(std::basic_string), hash_value(StringRef), and therefore
DenseMap<StringRef, X> lookups, with a flatten-and-call into
xxh3_64bits, a modern hash superior to CityHash.

hash_value(int) / hash_value(ptr) keep the existing Murmur-style
hash_16_bytes mixer; those are the dominant DenseMap key paths and a
fully-inline 16-byte mix beats inlining xxh3's larger 0..16-byte short
path.

To break dependency cycle: xxHash64, xxh3_64bits, and xxh3_128bits
ArrayRef/StringRef overloads move from llvm/Support/xxhash.h to inline
overloads in llvm/ADT/ArrayRef.h and llvm/ADT/StringRef.h, so xxhash.h
has no ADT dependencies.

A variant that inlined xxh3's 0..16-byte fast path at every
combine_bytes call site (vs. always calling out-of-line xxh3_64bits)
showed no measurable compile-time improvement on the tracker, so
combine_bytes is a one-liner over the out-of-line entry point.

llvm-compile-time-tracker.com (CTMark, instructions:u)
```
  stage1-O0-g           -1.76%   (sqlite3 -3.78%)
  stage1-aarch64-O0-g   -1.40%   (sqlite3 -2.86%)
  stage1-ReleaseLTO-g   -1.13%
  stage1-ReleaseThinLTO -0.45%
  stage1-O3             -0.43%
  stage1-aarch64-O3     -0.42%
  stage2-O0-g           -0.42%
  stage2-O3             -0.15%
  clang build           -0.71%   (wall -0.42%)
```

DenseMap-of-pointer paths (dominant at -O3) are untouched, so higher-
optimization configs see smaller wins as expected. opt's .text shrinks
~92 KB. Subsumes the StringRef-only carve-out proposed in #191115.

Notes on properties not introduced by this patch:

- Endianness: hash_combine over native integers was already not
cross-host
  stable. memcpy of a native integer into the buffer is host-encoded;
  fetch32 normalized the read but not the underlying bytes, so on LE vs
  BE the value fed to the mixer already differed. xxh3 inherits the same
  property: same byte stream, different mixer.

- Process seed: combine_bytes XORs get_execution_seed into the result,
  which cancels under hash_combine(x) ^ hash_combine(y). The pre-patch
  short/state paths fed the seed through hash_16_bytes / shift_mix
  non-linearly, so this is a regression in seed effectiveness under that
  pattern. Default seed is constant, so this only matters under
  LLVM_ENABLE_ABI_BREAKING_CHECKS. Follow-up: add a seeded xxh3 entry
  point in libSupport.

Aided by Claude opus 4.7