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

[Mehdi Amini via llvm-commits]

joker-eph wrote:

It seems to me like this is somehow a partial support for grid_constant, but instead of applying generally to any byval parameter, the support it is hard-coded for a special case (TMA). 
The general solution is likely not out-of-reach: likely something like applying the lowering logic in NVPTXLowerArgs.cpp to just any byval struct identified with the right property (as documented in NVVM IR doc here: https://docs.nvidia.com/cuda/nvvm-ir-spec/#supported-properties )

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