[llvm] [mlir] [MLIR][AMDGPU] Adding dynamic size check to avoid subword buffer load (PR #135014)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

First, I'm not convinced this sort of dynamic control flow is the way to go here at all and I want to have a broader discussion of how we got here 

*If* we're implementing a MLIR-side fix here - which I'm still not convinced of, especially since some of these tests are easier to inplement at the LLVM level, we need:
1. This only applies to sub-32-bit element types (might already be a thing, not 100% sure)
2. The wccesses has to be aligned to less than 4 bytes or the number of elements on the buffer isn't known to be a multiple of 4

https://github.com/llvm/llvm-project/pull/135014