Fix GeluFusion for unsupported CPU dtypes (#32431)
### Description
Fixes #32386.
For opsets below 20, `GeluFusion` replaces the standard Erf-based GELU
decomposition with `com.microsoft::Gelu(1)`. On CPU that contrib kernel
is registered only for `tensor(float)`, but the fusion currently accepts
`tensor(float16)` and `tensor(double)`. On Linux ARM64 an fp16 graph can
therefore become a fused node with no compatible CPU kernel and session
initialization fails.
This change makes `GeluFusion` validate the CPU dtype against the
operator domain it will emit:
- `com.microsoft::Gelu` CPU fusion is limited to `float`.
- ONNX `Gelu(20)` CPU fusion continues to support `float` and `float16`.
- Other execution providers retain their existing behavior.
### Tests
Adds graph-level regression coverage confirming that an opset-17 CPU
float graph still fuses while the same CPU float16 graph remains as the
valid standard ONNX decomposition.
---------
Signed-off-by: Sylvester Kaczmarek <sylvesterkaczmarek@users.noreply.github.com>
Co-authored-by: Sylvester Kaczmarek <sylvesterkaczmarek@users.noreply.github.com>
