[Subview] unify lowering on parent-shape + valid dims (PR344 follow-up)

[TaoTao-real]

Summary

This PR is a follow-up on top of #344 and keeps the implementation on the same line.
It unifies subview lowering semantics and removes compact/non-compact bifurcation.

What changed

• Remove compact/non-compact subview split and related attrs.
• Keep subview result tile shape as parent shape.
• Represent subview logical size via valid_row/valid_col (default from sizes, explicit valid clamped by size).
• Mark subview-origin bind with pto.view_semantics = "subview".
• In EmitC, handle subview-origin tile lowering separately from generic alloc-origin tile rebinding.

Files

• lib/PTO/Transforms/PTOViewToMemref.cpp
• lib/PTO/Transforms/PTOToEmitC.cpp
• lib/PTO/IR/PTO.cpp
• include/PTO/IR/PTOOps.td
• docs/PTO_IR_manual.md

Why

This aligns the implementation to the requested model:

• no compact/non-compact branch in analysis/lowering,
• preserve parent tile shape in type,
• use valid dims to describe sub-tile extent,
• keep codegen distinction by subview semantics instead of memory compactness tags.

Validation

Local

• runop.sh -t SubView: pass (with expected XFAIL)
• runop.sh -t Sync: pass (with expected SKIP/XFAIL)

Board (Ascend910, device 2, CI-like params)

• Command family: test/npu_validation/scripts/run_remote_npu_validation.sh
• Params: STAGE=run RUN_MODE=npu SOC_VERSION=Ascend910 DEVICE_ID=2
• PTO_ISA_COMMIT=893e4b00a826231e776539512b5b17f82fc11838
• SKIP_CASES=mix_kernel,vadd_validshape,vadd_validshape_dynamic,print,storefp,partition5d_a5,partition5d_dynamic_a5,mrgsort_a5,tmatmulk_autosync_a5
• Result: OK=166 FAIL=0 SKIP=9

[TaoTao-real]

Superseded by updating PR #344 directly. Closing this duplicate follow-up PR.