Refactor CPU ScatterAxis and MaskedScatter dtype dispatch

[PhysicistJohn]

Summary

Replace the remaining exhaustive payload dtype switches in CPU ScatterAxis
and MaskedScatter with the existing dispatch_all_types helper. This keeps
the restricted integer index dispatch, reduction mode selection, broadcasting,
source traversal, and diagnostics unchanged while removing 83 net lines.

Testing

• Release CPU-only build with global -Werror
• Native C++ suite (247/247 cases, 3,326/3,326 assertions)
• python/tests/test_ops.py (153/153 tests)
• focused vmap and autograd coverage for both operations
• 1,456 ScatterAxis parity cases and 14 exact unsupported-index diagnostics
• 602 MaskedScatter parity cases and 14 exact short-source diagnostics
• baseline/candidate symbol and linked-export sets match; the indexing object is 416 bytes smaller
• clang-format and git diff --check

[zcbenz]

Would you mind updating all the dispatch code in the file in one pull request? Our CI is throttled to handle many pull requests.

[PhysicistJohn]

Thanks, Cheng — agreed. I folded the remaining MaskedScatter payload switch
into this PR, so the last manual exhaustive payload-dtype router in
indexing.cpp now uses dispatch_all_types. The restricted integral-index
routers and reduction-mode switches remain unchanged because they encode
narrower type and operation semantics, including the existing diagnostics. The
combined diff is +8/-91. The Release -Werror build, 247/247 native cases,
153/153 Python ops tests, and paired ScatterAxis/MaskedScatter parity and
diagnostic checks all pass.