[llvm] [AMDGPU] Remove explicit PartialThreshold setting in loop unrolling (PR #198901)

[Ravil Dorozhinskii via llvm-commits]

ravil-mobile wrote:

@adelejjeh

The change resulted in an aggressive transformation (i.e.,  loop unrolling in a software pipelined loop) that led to increased register pressure beyond the physical register file capacity.

The change also broke the numerical correctness of a Triton-GEMM kernel by allowing aggressive loop unrolling that:
1. Destroyed the MFMA accumulator chain
2. Used wrong base accumulators for iteration 1
3. Failed to properly chain spill/reload operations

https://github.com/llvm/llvm-project/pull/198901