[llvm] [NVPTX] Add NVPTX intrinsics for TMA copies (PR #95289)

[Adam Paszke via llvm-commits]

apaszke wrote:

`grid_constant` is a new custom NVIDIA extension and in my view you're welcome to upstream it as well, but I don't see why this should block a PR that improves the current implementation. Especially that it's the only way to sensibly pass TMA descriptors as kernel args which is a pattern you, as a company, recommend. Happy to close this if you think the upstreaming will happen soon, but it's blocking for us and I can't wait indefinitely.

https://github.com/llvm/llvm-project/pull/95289