[CUDA] Add GatedAdd contrib operator (#31835)
## Description
Add `com.microsoft::GatedAdd`, a CUDA contrib operator that computes `X
+ round_to_T(Y * gate)` with a per-row gate broadcast across the hidden
dimension. The FP16 and BF16 implementations use separate PTX multiply
and add instructions to preserve the rounding behavior of separate ONNX
`Mul` and `Add` nodes rather than contracting to FMA semantics.
The operator is intended to remove one launch per gated residual/add
pair in MoE shared-expert paths. On Qwen3.6-35B-A3B-NVFP4 with N=3 MTP,
fusing 40 main-model pairs plus one MTP pair reduced median decode
latency from 7.311 to 7.225 ms/round (-1.18%). A graph-off Nsight
capture measured 40.35 fewer launches/round and 1.30% lower GPU kernel
time.
## Changes
- Add the `GatedAdd` schema for float, float16, and bfloat16.
- Add CUDA kernels with strict reduced-precision rounding semantics.
- Register all three CUDA kernel types.
- Add float, FP16, and BF16 contrib-op tests.
## Validation
- Runtime micro-model outputs are bit-exact with separate `Mul` + `Add`
for float, FP16, and BF16.
- `onnxruntime` and `onnxruntime_providers_cuda` targets build
successfully.
- `clang-format --dry-run --Werror` passes for all touched C++/CUDA
files.
- Real Qwen graph validation replaced exactly 40 main and one MTP `Mul`
+ `Add` pairs; all other nodes, initializers, inputs, and outputs
remained unchanged after normalizing fused edge names.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
