[llvm] [AMDGPU][wmma] - Add tied wmma intrinsic (PR #69903)

[Jessica Del via llvm-commits]

OutOfCache wrote:

> I think you should avoid breaking up the underlying opcode name in the intrinsic with tied. Is there a better suffix you could use?

I don't know about a better suffix, but I could move the 'tied' to the very end (`amdgcn_wmma_f16_16x16x16_f16_tied`). Would that clear up any confusion? 
I initially placed the suffix there because it sounds like an modifier of the accumulator matrices, i.e., there is the intrinsic with any f16 accumulators, and there is the intrinsic with _tied_ f16 accumulators. Moving the suffix to the end still sounds reasonable, then it becomes a modifier of the entire intrinsic itself.

https://github.com/llvm/llvm-project/pull/69903