fix(hip): unify broadcast and reduction shape contracts

[fhanuman]

Summary

Unify NumPy broadcast and ONNX reduction result-shape rules across converter destination construction, HIP reification, and static verification.

• use one right-aligned broadcast rule for elementwise operations and converter destinations;
• preserve legal zero/unit broadcasting with select(lhs == 1, rhs, lhs) rather than integer maximum;
• share variadic Max/Min lowering through one pairwise broadcast chain;
• centralize axes, keepdims, and noop_with_empty_axes handling across all six reducers, including ReduceL2;
• use one output-to-input dimension map so keepdims = 0 preserves the correct surviving dimensions;
• preserve the validated outs fallback when reduction axes are only known at runtime.

Stack

This is PR 5 of 12 and depends on PR 4:

1. Constant carrier/externalization
2. Shape-rule foundation
3. Reshape provenance
4. MatMul/Gemm
5. Broadcast/Reduction ← this PR
6. Loop carriers
7. Payload/Readback
8. Conv/Pool
9. Gather/Tensor
10. Attention/Normalization
11. Explicit contract enforcement
12. Refinement hardening

AI assistance

AI tools assisted with the shape-rule consolidation and stack reconstruction. The broadcast and reduction layer was checked with the authoritative local build and focused rank-zero, keepdims, noop, converter, reification, and verifier tests.

Made with Cursor

[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	34.957	368640	0

Threshold: 0.01 | Run: 3952 - Commit: 354c3f0

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.26	361	3	1243
GroupQueryAttention_seq128	4359.05	1.75973	12	6	313
matmul_down_seq128	523.29	2.41	71	3	354

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.59	46.67	359	3	15589

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	failed	-	-	-	-

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	182.0	79.4	1.33	13.53
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	335.5	40.1	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	208	39.6

Run: 3952 - Commit: 354c3f0