Fix GEMM profiler verification across all batches

[meiyou666]

Summary

• Compare each logical GEMM batch independently through batch_data() and batch_stride(), while preserving the existing bitwise-exact and epsilon-based relative comparison semantics.
• Apply batched comparison to standard GEMM host, device, and cuBLAS verification; block-scaled GEMM D and SFD outputs; and blockwise GEMM D output.
• Set CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT and CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET on the cuBLASLt A/B/C/D layouts for batched GEMM.
• Limit the generic reference loop to host and device reference providers so dedicated provider results remain intact.

Validation

• Completed full profiler builds on an RTX 4090 with CUDA 11.8 using both CUTLASS_ENABLE_CUBLAS=OFF and CUTLASS_ENABLE_CUBLAS=ON.
• Ran 38 total host-reference matrix cases across the baseline and fixed profilers at beta=0, using batch_count values 1, 2, 3, 7, and 16. Clean cases passed. On the baseline, batch-0 faults were detected while nonzero-batch faults reproduced the false pass. With the fix, every injected fault was detected.
• Ran 19 cuBLAS matrix cases with the fixed profiler at beta=0 across the same batch counts and fault positions. Clean cases passed, and all batch-0 and nonzero-batch faults were detected.
• The full builds compiled the block-scaled and blockwise paths; those paths were not run on this GPU.

Fixes#3511.