Verify each finding against the current code and only fix it if needed.

Inline comments:
In
`@tensorrt_llm/_torch/cute_dsl_kernels/blackwell/blockscaled_contiguous_grouped_gemm.py`:
- Around line 87-90: Initialize problem_shape_ncluster_mnl before the swizzle
branch by computing it with cute.ceil_div(...) (using the existing problem_shape
and cluster sizes) and assign that value to a local variable
problem_shape_ncluster_mnl; then in the existing swizzle branch (where
swizzle_size > 1) call cute.round_up(...) against that precomputed
problem_shape_ncluster_mnl (instead of referencing
self.problem_layout_ncluster_mnl.shape which is not yet set), and finally assign
self.problem_layout_ncluster_mnl from the finalized shape after swizzling so
later code can safely use self.problem_layout_ncluster_mnl; update references to
raster_along_m and swizzle_size as needed but keep the ceil_div-based
initialization outside the if-block.
- Around line 63-68: The hooked_PersistentTileSchedulerParams_init currently
reads self.problem_layout_ncluster_mnl.shape before that attribute is guaranteed
to be assigned (causing AttributeError when swizzle_size > 1) and also applies
process-wide CUTLASS global mutations duplicated across three kernel modules; to
fix, ensure problem_layout_ncluster_mnl is initialized before any access in
hooked_PersistentTileSchedulerParams_init (assign a default/layout
unconditionally or move the shape access after the conditional assignments so
both swizzle and non-swizzle paths set self.problem_layout_ncluster_mnl), and
consolidate the global scheduler monkey-patch (the mutations currently
duplicated in blockscaled_contiguous_grouped_gemm.py,
blockscaled_contiguous_grouped_gemm_finalize_fusion.py, and
blockscaled_contiguous_gather_grouped_gemm_swiglu_fusion.py) into a single
module-level patch point to avoid import-order dependent behavior (remove the
duplicate overrides and reference that single centralized patch).

---

Nitpick comments:
In
`@tests/scripts/cute_dsl_kernels/run_blockscaled_contiguous_grouped_gemm_finalize_fusion.py`:
- Around line 1074-1097: The call to run(...) uses a long positional tail for
runtime flags which risks misbinding; change the final arguments to keyword
arguments for clarity and safety by passing permuted_m=permuted_m, topk=topk,
seq_len=seq_len, raster_along_m=raster_along_m, use_blkred=use_blkred, and
use_cupti=use_cupti (and any other trailing booleans) so the invocation of the
run function uses explicit parameter names and won't break if the signature
changes.