[Feature Request] Improved pattern matcher for AR+residual_add+RMSNorm fusion (cast/reshape coverage)

Summary

The AutoDeploy fuse_allreduce_residual_rmsnorm transform only matches the exact chain all_reduce -> add(residual) -> rmsnorm. On real FP8 / MoE models the exported graph interposes either a shape-only reshape/view (between the all-reduce and the residual add) or a dtype aten.to.dtype cast (between the residual add and the RMSNorm). The matcher then silently misses the site and the production fused AR kernel is never used.

Measured impact

• nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 (TP=4): 52 trtllm_dist_all_reduce nodes, fusion matched 1/52 — the residual stream is carried through add -> aten.to.dtype(bf16) -> rmsnorm.
• Qwen/Qwen3.5-35B-A3B-FP8 (TP=4): 80 all-reduce sites, matched 40/80 — the 40 MLP/MoE sites are all_reduce -> aten.reshape -> add -> rmsnorm.

This disables fusion for both the trtllm and the experimental cute backends, since they share this matcher.

Proposed approach

• Cast variant: add all_reduce -> add -> aten.to.dtype -> rmsnorm pattern(s), rewiring the fused op's residual (second) output onto the cast node. The fused kernels already emit a bf16 residual, so the cast is absorbed. The matcher util itself flags aten.to.dtype node elision (utils/pattern_matcher.py, register_ad_pattern docstring point 3c) — handle via a dedicated pattern + op_ignore_types on the cast dtype arg.
• Reshape variant: tolerate a shape-only aten.reshape/view between the all-reduce and the residual add; fold it into the fused op (ignore the shape literals via op_ignore_types).
• Keep existing variants; parameterize the factory over {add_order} x {rmsnorm_op} x {cast?} x {reshape?}.

Scope

• Surface: tensorrt_llm/_torch/auto_deploy/transform/library/collectives.py
• Tests: tests/unittest/auto_deploy/.../test_allreduce_residual_rmsnorm_fusion.py

Tests / validation

• Graph-level match-count unit test (CPU/meta): synthetic modules with an intervening cast and reshape; assert the fused op appears and num_matches increases.
• Extend the existing multi-GPU correctness test with the cast and reshape cases.

Risk

library-visible (matcher shared by the trtllm + cute fusion backends; AutoDeploy multi-GPU CI stages required). Not API-visible.

Links

• Related PR: (filled in once the PR is opened)
• Discovered during evaluation of PR #14534 (experimental CuTe fused AR+residual+RMSNorm kernel).
• JIRA: None