Fix module-level capability gates in test_kernels.py and
test_fp8_grouped_mm.py that used raw get_device_capability() >= (9, 0),
which incorrectly passes on MI250X (gfx90a reports 9.0 but lacks FP8).
Replace with is_sm_at_least_90() || is_MI300() || is_MI350().

test_kernels.py:
- Enable 5 FP8 kernel tests on MI300/MI350: jagged rowwise scales,
  jagged rowwise scales via transpose, jagged colwise scales,
  3D rowwise transpose-rhs (atomic), and 3D rowwise transpose-rhs
  (reduction). These kernels already have AMD-specific autotune configs
  on main.
- Use e4m3_dtype from config instead of hardcoded torch.float8_e4m3fn
  (MI300 uses float8_e4m3fnuz).
- MXFP8 tests remain skipped on ROCm (MXFP8 format not yet supported).

test_fp8_grouped_mm.py:
- Fix module-level gate so the file loads on FP8-capable ROCm hardware.
  The main test (test_fp8_rowwise_scaled_grouped_mm) already has proper
  per-test MI300/MI350 gating and ROCm-specific tolerances.
- Enable test_K_or_N_dim_not_multiple_of_16 on ROCm (validates dimension
  assertions, uses Float8TrainingOpConfig which already resolves to the
  correct platform FP8 dtype).

test_fp8_grouped_mm.py:
- Merge duplicate `from torchao.utils import` statements (I001)
- Remove unused `skip_if_rocm` import (F401)

test_kernels.py:
- Sort imports alphabetically so torchao.float8 precedes
  torchao.prototype (I001)

…ernels.py

Apply `ruff format` to wrap long pytest.skip() calls and multi-condition
if-statements that exceed the line length limit.