Enable NVFP4 fused grouped MLP SwiGLU

[sraman-rgb]

Stacked on NVIDIA#2971.

This PR adds the incremental NVFP4 grouped MLP SwiGLU fusion changes on top of PR 2971, without modifying PR 2971 itself and without the temporary Python bulk_allocate fallback.

Changes:

• Enable NVFP4 recipes in grouped MLP fusion selection.
• Add NVFP4 forward/backward fused grouped MLP SwiGLU handling.
• Use grouped tensor GEMM for NVFP4 FC1 dgrad with grouped_fc1_weight directly.
• Fix NVFP4 discrete-input grouped GEMM layout metadata.
• Avoid the single-group split-size host sync in grouped linear.

Validation:

• NVTE_GROUPED_LINEAR_SINGLE_PARAM=1 NVTE_CUTEDSL_FUSED_GROUPED_MLP=1 python3 -m pytest -q --tb=short -ra tests/pytorch/test_fusible_ops.py::TestSequentialModules::test_grouped_mlp -k 'scaled_swiglu and nvfp4' -> 48 passed, 336 skipped
• NVTE_GROUPED_LINEAR_SINGLE_PARAM=1 NVTE_CUTEDSL_FUSED_GROUPED_MLP=1 python3 -m pytest -q --tb=short -ra tests/pytorch/test_fusible_ops.py::TestSequentialModules::test_grouped_mlp -> 1216 passed, 4544 skipped