fuse_attention: exclude side-input reductions from the decode (kv_cache) attention group

[rlegithub]

Summary

find_kv_cache_attention::get_attn_instructions walks attention inputs backward and pulls valid side-input ops into the attention submodule. Because reduce_* ops are valid attention ops (softmax), it also pulled a preceding RMSNorm's reduce_sum into the group. That reduction is an intermediate feeding the attention GEMM, and rocMLIR's TosaToRock reduction lowering can only lower a reduce whose output traces to a kernel result → tosa.reduce_sum can't trace the reduction output to a kernel result → get_tuning_config throws → "No valid tuned compilation".

Fix: skip reduce_* ops in the backward side-input expand. The attention's own softmax reduces are on the direct start→end path (find_instructions_between) and are unaffected, so the attention stays fused; only the RMSNorm reduce is left as its own kernel. A reduce feeding a GEMM can never trace to a kernel result, so excluding it cannot regress a working fusion.

This restores fused (flash) decode attention for models with a pre-attention RMSNorm on gfx1151.

Note: this is a pragmatic fusion-boundary fix. The more complete solution — teaching rocMLIR to fuse an intermediate reduction into the GEMM prologue (norm + attention as one kernel) — is a larger rocMLIR change tracked separately.

Test plan

• CI
• Verified on gfx1151: int4 LLM (Llama-4-family) decodes coherently with fused attention; Llama-3.2-1B int4 regression passes.

[pfultz2]

Unit tests need to be added for this.

[pfultz2]

Unit test needs to be added and the comments cleaned up.