fix(use-output-allocator): report graph-output rank for reshaped-view outputs

[anilmartha]

Summary

UseOutputAllocator rewrote a graph-output memref.alloc into hip.alloc_output using the producer alloc's type. When the returned value is a reshaping view of that alloc (e.g. a rank-2 Gemm result expand_shape'd to rank-3 with no trailing op — DETR's class_labels_classifier logits head), the EP registered the alloc's rank with ORT instead of the declared graph-output rank, so ORT rejected the output:

Invalid rank for output: logits Got: 2 Expected: 3

Outputs with a trailing op (e.g. pred_boxes' sigmoid, whose alloc is already at the output rank) were unaffected — which is why only logits failed.

Fix

• Allocate hip.alloc_output at the func.return operand type (the graph-output shape).
• When it differs from the producer alloc's type, rebuild the inverse of the view chain (cast / expand_shape / collapse_shape, static output) so the producer keeps writing into the same EP-owned buffer viewed as its original type — no extra copy.
• Fast path (return type == alloc type) is unchanged; subview or dynamic-shaped outputs fall back to the legacy behavior, so no existing case regresses.

Applies to both the LLVM_IR and NATIVE artifact paths.

Test plan

• facebook/detr-resnet-50 (fp16): logits now returns {1,100,92}; onnxruntime_perf_test runs clean.
• Re-run existing lit/numeric suites for regressions (esp. conv/reshape output models).
• Spot-check a model with a rank-4 conv output collapsed before return.

[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.

[amd-mingw]

#533 is fixing on it. cc @amd-genmingz

[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: 3271 - Commit: c56b74c

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.53	6.39	360	3	1244
GroupQueryAttention_seq128	4377.53	1.65253	12	6	310
matmul_down_seq128	524.48	2.34	75	3	352

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.58	44.35	369	3	15589

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.56	9.07	359	3	15759

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	164.7	80.2	1.33	13.57
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	282.2	39.6	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	193	39.2

Run: 3271 - Commit: c56b74c