Fix lowering of pow() in D3D12Compute when the base is provably positive

[shoaibkamil]

We had an internal report of a codegen error when lowering pow() in the D3D12Compute backend, producing incorrect HLSL like the following:

pow(_183, float _367 = _364 * _364;

The root issue was that this lowering path was directly injecting into the string representing the shader, rather than using a lowering path that will handle dealing with the arguments correctly. Also added a test that failed before this change; I can separate it out to its own file if that's better.

[slomp]

Thanks for spotting the issue and fixing it!

[mcourteaux]

Glancing at this PR, I wonder why the original code was bad. Why doesn't print_expr() print an Expr?

[alexreinking]

Glancing at this PR, I wonder why the original code was bad. Why doesn't print_expr() print an Expr?

Apparently it doesn't return an rvalue expression in all cases. It ought to, imo, but that would be out of scope for this PR.

[alexreinking]

Getting some incorrect values on webgpu -- that must be an llvm_main injection.

[mcourteaux]

@alexreinking I think we can land this?
Yes. Landed it. We have a bit of a backlog with good PRs that didn't get merged because of LLVM shenanigans.