[devinhuang2001]

Fixes #2112

MatMul8bitLt.backward scales row-quantized grad_output with column statistics. Use the column-quantized tensor that matches those statistics.

The CUDA path also needs CAt.t() materialized contiguously: its int8 GEMM passes raw pointers and shape-derived leading dimensions without honoring transpose strides. On a T4, correcting only the quantization pair still gives a 1.47 relative weight-gradient error; making the second operand contiguous reduces the single-operator error to 0.0090.

Adds 16 regression cases covering 2D/3D inputs, thresholds 0/6, input gradients enabled/disabled, and both bnb.matmul and Linear8bitLt. The oracle computes grad_output.T @ input in fp32 with deliberately different gradient-column scales.

Validation on Tesla T4, PyTorch 2.11.0+cu130, source baseline a616886 (using the packaged CUDA 13.0 native library):

• New regression cases: 16 failed on baseline; 16 passed with both corrections.
• 32 selected existing CUDA test_matmullt cases: passed (selected across the collected case list).
• Full pre-commit run --all-files: all 10 hooks passed.
• git diff --check: passed.

The full test suite and other accelerator backends were not run.