Fix QuickGelu fusion for unsupported CPU dtypes (#32427)
### Description
Fixes #32421.
`QuickGeluFusion` can replace a valid CPU `Mul -> Sigmoid -> Mul`
QuickGELU decomposition with `com.microsoft::QuickGelu` for
`tensor(double)`. The CPU QuickGelu kernel is registered only for
`tensor(float)`, so enabling extended graph optimizations can turn a
runnable model into one that fails during session initialization.
The fusion now checks the execution provider and only creates QuickGelu
for CPU inputs supported by the CPU kernel. Other execution providers
keep their existing behavior; in particular, CUDA continues to support
`double`.
Regression coverage verifies that CPU `float` graphs are still fused and
CPU `double` graphs remain unfused.
### Motivation and Context
Graph optimizations must preserve model executability. This keeps the
standard ONNX decomposition when fusing it would create a contrib op for
which the selected CPU execution provider has no compatible kernel.
