[MLIR][gpu-to-llvm] Support multiple async dependencies when calling gpu.launch_func

[ianayl]

Adds a new pattern to gpu-to-llvm pass that generates additional gpu.waits before gpu.launch_func in the case that gpu.launch_func has multiple async dependencies. Without this PR, gpu.launch_func ops would fail to convert/legalize to GPU runtime calls when multiple async dependencies are provided.

Notes/Request for comment:

• This PR was written in a way such that support for other operations that also require a single async dependency, i.e. gpu.alloc can also benefit from the pattern and implicitly support multiple async dependencies. The question is, would people like this in MLIR?
• Additionally, should this be an option instead of enabled by default?

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.