DISABLED test_main_loop_scaling_shape1_use_fast_accum_True_scaling_block_sizes2_cuda (__main__.TestFP8LoweringCUDA)

This test fails consistently on CUDA (seen on `Limited CI on H100 / linux-jammy-cuda13.0-py3.10-gcc11-sm90 / test-osdc (smoke, 1, 1, mt-l-x86iamx-22-225-h100)`) with:

```
ValueError: scale_b must have shape N x ceil_div(K,128) Float elements, got [transposed]
```

Root cause: for `w = BlockWise1x128` recipes, the test builds `scale_b` (w's inverse scale) in `[K/128, N]` layout, but eager `torch._scaled_mm_v2` requires `[N, K/128]`. This is a pre-existing test/kernel scale-layout mismatch; CI only began exercising it on 2026-06-29.

Repro:
```
python test/inductor/test_fp8.py TestFP8LoweringCUDA.test_main_loop_scaling_shape1_use_fast_accum_True_scaling_block_sizes2_cuda
```

One of ~14 failing `test_main_loop_scaling` parametrizations (`w = BlockWise1x128` recipes). See also #188675, #188704.

cc @mruberry @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @aakhundov @coconutruben @jataylo

Start by running the reproducer in test/inductor/test_fp8.py for TestFP8LoweringCUDA.test_main_loop_scaling_shape1_use_fast_accum_True_scaling_block_sizes2_cuda. Inspect how scale_b is built for the BlockWise1x128 recipes and compare it with the eager torch._scaled_mm_v2 shape requirement. Done means the affected test_main_loop_scaling parametrizations run successfully on CUDA without being disabled.

Written by the indexing model from the issue text.