Support opset 28 SwiGLU and log reductions

[Justin Chu (justinchuby)]

Summary

• add a CPU float/double kernel for ONNX SwiGLU-28, including the vectorized MLAS SiLU path for the default alpha
• validate the upstream function body can be loaded and explicitly inlined, plus shape inference and numerical behavior
• split ReduceLogSum and ReduceLogSumExp registrations at opset 28 across CPU, CUDA, JS, and WebGPU so v28 advertises only supported float types while opsets 18-27 retain integer support
• update generated operator registration documentation

The native SwiGLU kernel is narrowly required because the upstream body expands to Swish-24, which ORT cannot safely re-expand at an outer opset 28 after referenced primitive schemas changed; without the kernel, session initialization fails with no Swish(24) implementation.

Validation

• Windows RelWithDebInfo onnxruntime_test_all build
• Windows RelWithDebInfo onnxruntime_provider_test build
• FunctionTest.SwiGLUOpset28*: 2 passed
• ReductionOpTest.ReduceLogOps_Opset*: 4 passed
• clang-format and git diff --check
• independent GPT-5.6 Terra review: no remaining findings

Stacked on #32359 / justinchuby-onnx-1-23-opset-28-base at 66026459a8a7553a04dfcb5798ece21f9a4ba0b1.