Remove Cast before and after Gelu

[Lafi7e]

Some mix-precision models will cast data to float before Gelu and cast back to half after. This PR changes the Gelu and GeluGrad CUDA kernel to use float for compute internally, and remove these Cast nodes from graph on CUDA EP.

• Change Gelu/GeluGrad CUDA kernel to use float internally will not harm the perf, while we can have better precision.
• This PR doesn't handle ROCm EP as currently we don't have Gelu/GeluGrad ROCm kernel.
• This PR doesn't handle FastGelu, as the kernel implementation for FastGelu is to complicated to changed to use float internally.

[tianleiwu]

Is there a way to generalize the fusion for element-wise operator, like
x-->Element-wise Op1 --> Element-wise Op2--> can be fused as Element-wise Op2(Op1(x))

[ytaous]

Wouldn't the Cast Propagation feature (#7454) that Satya implemented already covers this case?

[Lafi7e]

Wouldn't the Cast Propagation feature (#7454) that Satya implemented already covers this case?

Right. Previously I just want to enabled this for CUDA only, but the PropagateCastOps cannot do this. But on second thought, we don't have MLFloat16 kernel for CPU for Gelu/GeluGrad, it's safe to use PropagateCastOps. I've changed the code to use PropagateCastOps by promoting Gelu from Level2 Ops to Level1, which also solves this issue.