Refactor CPU Softmax dtype dispatch

[PhysicistJohn]

Summary

Replace CPU Softmax's explicit dtype switch with dispatch_all_types.
The compile-time branches preserve the existing accumulator policy: float16
and bfloat16 use float accumulation in precise mode and their native types
otherwise, while float32 and float64 keep native accumulation. Other types
retain the existing unsupported-type error.

This removes eight net lines from one function.

Testing

• Release CPU-only build with C++20 and warnings as errors
• Baseline/candidate raw-byte parity across float16, bfloat16, float32, and
  float64, with precise mode both enabled and disabled (32 cases; transcripts
  byte-identical)
• Python operations suite (150/150 tests)
• Native C++ suite (247/247 cases, 3,326/3,326 assertions)
• Global object symbols preserved
• Composition check with the older same-file Add support for x86 SIMD with Highway #3019: both patch orders have
  only the adjacent include insertion to resolve; retaining both includes
  produces identical source in either order, the combined Release library
  builds, and the focused reduction/softmax suite passes
• clang-format and git diff --check