fix(layout): correct row reduction layout inference for rowmin/rowmax/rowsum

[TaoTao-real]

Summary

• fix row-reduction layout propagation in InferPTOLayout so rowmin/rowmax/rowsum keep layout-consistent tile/global tensor combinations
• preserve key tload/tstore attribute information across the layout/view lowering path
• tighten consumer-preference inference to only apply in safe ambiguous vector cases (avoid over-forcing layout)

Scope

This PR is intentionally scoped to row-layout inference/lowering behavior and does not include the AllocToPointerCast change already tracked separately.

Files

• lib/PTO/Transforms/InferPTOLayout.cpp
• lib/PTO/Transforms/PTOViewToMemref.cpp
• lib/PTO/Transforms/PTOToEmitC.cpp
• test/samples/runop.sh

Validation

Local generation

• runop --enablebc -t Rowmin
• runop --enablebc -t Rowmax
• runop --enablebc -t Rowsum
• all generated successfully

Remote board (A3, CI-like script)

• RUN_ONLY_CASES=rowmin,rowmax,rowsum
• result: OK=3 FAIL=0 SKIP=0
• all three cases built + ran + compare passed

Artifacts (local):

• /private/tmp/rowlayout_pr_clean_3cases_20260307_145409/remote_npu_validation_results.tsv
• /private/tmp/rowlayout_pr_clean_3cases_20260307_145409/remote_run.log

Risk

• behavior change is limited to layout inference/lowering decisions
• no InsertSync/event-allocation logic change in this PR