[InsertSync] scalar autosync + addptr alias-through (scalar-only scope)

[TaoTao-real]

Summary

• make pto.load_scalar / pto.store_scalar participate in InsertSync as PIPE_S ops
• add pto.addptr alias-through in InsertSync translator, including constant-offset byte folding and conservative handling for dynamic offsets
• keep scalar sync behavior aligned with AscendNPU-IR semantics:
  • skip PIPE_S -> PIPE_S local sync insertion
  • keep scalar cross-pipe dependencies as normal set_flag/wait_flag sync (no forced PIPE_ALL fallback)
• add scalar sync regressions and runop structural checks:
  • test_inject_sync_scalar_cross_pipe.py
  • test_inject_sync_scalar_intra_pipe_barrier.py

Scope

This PR is intentionally scalar-sync only.
Layout-inference changes (e.g. rowmin/rowsum/rowmax related fixes) are explicitly excluded from this PR.

Motivation

InsertSync previously could miss synchronization around scalar load/store paths and addptr alias chains, which risks under-synchronization in scalar-memory-involved dependencies.

Design

• LoadScalarOp / StoreScalarOp
  • add OpPipeInterface
  • return PIPE_S in getPipe()
• InsertSync analysis
  • same-pipe PIPE_S is skipped (in-order scalar pipeline)
  • scalar cross-pipe dependencies are synchronized via set_flag/wait_flag
• alias propagation
  • add pto.addptr to alias chain updates
  • support constant addptr offsets in byte-range analysis
  • dynamic addptr offsets use conservative alias modeling (prefer over-sync to under-sync)

Testing

Local regression:

source /Users/lishengtao/Workspace/PTO/env.sh
export PTO_PYTHON_ROOT=/Users/lishengtao/Documents/PTO/PTOAS_upstream_main_main/build-main/python
pto_export_runtime
PTOAS_BIN=/Users/lishengtao/Documents/PTO/PTOAS_pr206_scalar_only/build-main/tools/ptoas/ptoas \
  bash test/samples/runop.sh -t Sync

Result: pass (A5-only sample skipped as expected).

A3 remote board validation (latest, commit 6c8c44e):

• Host: 101.245.68.6 (DEVICE_ID=10, SOC_VERSION=Ascend910)
• Flow: CI-equivalent test/npu_validation/scripts/run_remote_npu_validation.sh
• Result: OK=19, FAIL=0, SKIP=1
• Key scalar cases:
  • test_inject_sync_scalar_cross_pipe: OK
  • test_inject_sync_scalar_intra_pipe_barrier: OK

Risk / Rollback

• Risk: scalar-related sync coverage expands compared with previous scalar-untracked behavior
• Mitigation: same-pipe PIPE_S remains skipped; cross-pipe behavior follows existing set/wait model and has local + remote regression coverage
• Rollback: revert this PR (changes are localized to InsertSync + Sync tests)

[TaoTao-real]

Added commit 269dedb for the long-term row-reduction layout fix and validated on A3 board before PR submission.

A3 remote validation (CI script flow):

• Script: test/npu_validation/scripts/run_remote_npu_validation.sh
• Env: STAGE=run, RUN_MODE=npu, SOC_VERSION=Ascend910, DEVICE_ID=2
• Run only: rowmin,rowsum,rowmax

Result:

• rowmin: OK
• rowsum: OK
• rowmax: OK
• Summary: OK=3 FAIL=0

Extra note: when additionally running tensor_view_layout_dn, it fails with a pto-isa static assert (DN GlobalTensor with RowMajor MatTile) on A3; this case is orthogonal to the row-reduction fix above.