fix(matmul-lowering): use extractMemRefDataPtr for offset-correct pointers

[kirankumarmatam-amd]

Problem

MatmulOpLowering used extractContiguousMemRefPtr which returns only
alignedPtr from the memref descriptor, ignoring the offset field.

When onnx.Split produces Q/K/V slices via memref.subview, each slice
carries a non-zero offset into the parent buffer. extractContiguousMemRefPtr
returned the parent buffer's base address for all three, so Q and V MatMuls
silently read K's data instead of their own — reducing end-to-end model
accuracy to ~0.36 cosine on Qwen 7B VLM (gfx1151).

Fix

Switch A, B, and output to extractMemRefDataPtr, which computes
alignedPtr + GEP(offset) — the GEP advances by offset elements,
correctly accounting for element size:

// Before:
Value APtr = extractContiguousMemRefPtr(adaptor.getA(), rewriter, loc);
// After:
auto AType = cast<MemRefType>(op.getA().getType());
Value APtr = extractMemRefDataPtr(adaptor.getA(), AType,
                                  getTypeConverter(), rewriter, loc);

Testing

Verified on gfx1151 (Strix Halo, Windows D3D12 HMM): cosine similarity
for Qwen 7B VLM restored to ~1.0 after this fix.

[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: 3531 - Commit: 918c161

[github-actions]

MorphiZen EP Performance Results

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.51	6.29	425	3	1242
GroupQueryAttention_seq128	4284.47	1.70747	11	6	311
matmul_down_seq128	523.83	2.45	75	3	352

EPContext Export Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.55	44.50	364	3	15589

EPContext Import Performance

Model	QPS	Session (s)	1st Infer (ms)	CPU%	Mem (MB)
full_model_seq128	7.53	9.14	360	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	172.0	79.0	1.33	13.54
Llama-3.1-8B-awq-g128-int4-asym-fp16-onnx-dml	1	5	128	128	325.3	41.1	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	205	40.4

Run: 3531 - Commit: 918c161