[LLVMGPU] Add ROCDLLoadToTransposeLoadPass to TileAndFuse pipeline

[Max191]

Adds the ROCDLLoadToTransposeLoad pass to the LLVMGPUTileAndFuse pipeline. This is only enabled for ROCDL, and a test flag is added to turn off the feature if needed (mainly for benchmark testing).

Convolution Benchmark Results

The data below is all for bf16 convolutions. I didn't put comprehensive GEMM (MxK @ KxN) data in a spreadsheet, but for GEMM, the speedup is in the range of 0-17% for f16 and 0-60% for i8 GEMMs.

Full spreadsheet of results: https://docs.google.com/spreadsheets/d/1QEwemqviUzk4GginGdaDT7u8pP9x_Bku7r1aAvOUOCk/edit?usp=sharing

Weight Backward Convolutions

Equates to KxM @ KxN GEMM layout.

Benchmark Summary:
Total benchmarks: 162
Significant changes (>2.0%): 142
Improvements (transpose_load faster): 133
Regressions (default faster): 9
Mean % change: -20.44%
Range: -56.29% to 12.22%

Input Backward Convolutions

Equates to MxK @ KxN GEMM layout.

Benchmark Summary:
Total benchmarks: 146
Significant changes (>2.0%): 43
Improvements (transpose_load faster): 24
Regressions (default faster): 19
Mean % change: -1.09%
Range: -34.02% to 12.08%

Additional notes

• All benchmarks were collected on a single MI355 GPU
• There are "regressions, but the top regressions appear to be from noise. Retesting them did not show real regressions.
• Forward convolution data is not included, since forward convolutions do not target transpose_load instructions.

ci-extra: test_torch

[Max191]

Based on #23267

I haven't collected comprehensive benchmark numbers, so I will share them in the PR description once I have collected them.

[Max191]

I was finally able to get some good benchmark runs, and I added the results to the description.

cc @nirvedhmeshram @yzhang93 @MaheshRavishankar

[nirvedhmeshram]

LGTM, great to see performance improvement on weight backward!