feat(onnx-to-hip): fold attention score prep Mul and broadcast Add

[amd-mounikk]

Summary

• Add pre-lowering patterns for Swin-style window-attention score preparation before lowerOnnxConstants.
• FoldMatMulScaleMul: rewrites MatMul(A,B) -> Mul(scores, scale) to Mul(A, scale) -> MatMul(A', B), avoiding a broadcast mul on large score tensors.
• ExpandConstantBroadcastAdd: compile-time expands broadcast constant biases so Add(activation, Constant[bias]) becomes same-shape add, avoiding runtime broadcast elementwise on large attention maps.
• Review fixes (3bfb091): MatMul scale fold guarded by K/N extent check; bias expansion capped at 1 Mi output elements.
• Includes LIT coverage in test_attention_score_prep_opt.mlir (positive + negative cases for both patterns).

Test plan

• pre-commit run --from-ref origin/main --to-ref HEAD (lintrunner + licenseheaders)
• hip-mlir-opt test/lit/Conversion/onnx-to-hip/test_attention_score_prep_opt.mlir (FileCheck pass)
• SwinV2 tiny 8x521x1024 end-to-end perf (local): avg ~274 ms/inference, matmul 27 calls / 31.8 ms GPU

[github-actions]

Thanks for opening a PR!

This project follows LLVM's incremental-development and AI-tool-use
guidance. See CONTRIBUTING.md
for the project workflow.

Before requesting review, please check that:

1. The change is focused. Substantial work links the relevant issue
   or design discussion.
2. The PR documents relevant test results and updates affected
   documentation.
3. If AI tools provided substantial assistance, the description
   explains what was assisted and how it was validated, and commit
   trailers identify the tool. The contributor has reviewed and
   understands the result.

Reviewers are assigned through
CODEOWNERS where ownership
is configured.

[github-actions]

L2 Accuracy Results (EP vs CPU)

Model	Combined L2	Total Elems	Skipped NaN/Inf
conv_test_hybrid	4.8668E-07	64	0
GroupQueryAttention_seq256	25.2366	2621440	0
MatMulNBits_o_seq128	259.906	368640	0
QMoE_seq128	34.9552	368640	0

Threshold: 0.01 | Run: 3998 - Commit: d3aa112

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.50	6.36	360	3	1243
GroupQueryAttention_seq128	4559.40	1.71849	11	6	313
matmul_down_seq128	495.18	2.38	71	3	354

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.58	45.38	361	3	15589

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.52	9.80	358	3	15761

OGA Benchmark Results

Model	Warmup	Reps	Prompt Len	Gen Tokens	TTFT (ms)	TPS	Peak Mem (GB)	GPU Mem (GB)
gpt-oss-20b-webgpu-int4-rtn-block-32	1	5	128	128	180.9	80.5	1.33	13.53
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	272.3	40.2	1.22	6.43

OGA Wheel Smoke (Python benchmark_e2e.py)

Model	TTFT (ms)	TPS
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	192	39.4

Run: 3998 - Commit: d3aa112

[fhanuman]

I will approve the PR to unblock the merge. However, I am still not clear on how much this change improves performance, so in future PRs like this, it would be helpful to include an example showing where it helps and some before-and-after performance numbers.