[codex] Improve PTODSL frontend ergonomics and API coverage

[zhoubot]

Summary

• add constexpr specialization support, generated micro-op wrappers, and a more pythonic pto tensor/tile facade
• add a public API snapshot plus host-side API/regression coverage, and split CI into fast and full workflows
• expose the missing pto.barrier_sync helper and add sync-specific frontend/API coverage

Why

PTODSL was missing a coherent contract and regression layer around the frontend surface while the PTO frontend itself was growing in three directions at once: constexpr specialization, micro-op coverage, and a more ergonomic tensor/tile API. This PR packages those frontend changes together and adds the API hygiene gates needed to keep the public surface intentional.

What Changed

• frontend/runtime surface:
  • Constexpr, const_expr, and range_constexpr specialization flow
  • generated micro module plus pto re-exports from the PTOAS vPTO manifest
  • pythonic pto.ptr, make_tensor, TensorView.slice, make_tile_buffer, and tile-buffer wrappers
  • pto.barrier_sync wrapper added to the public sync surface
• coverage and tooling:
  • new tests/api contract suite with snapshot-backed public export checks
  • new tests/regression host regression suite
  • shared PTOAS manifest locator for local and CI use
  • snapshot updater script
• CI:
  • new fast host-only workflow for PRs
  • existing matrix workflow narrowed to full regression on main/manual runs

Validation

• PYTHONPATH=... python -m pytest -q tests/api -> 48 passed
• PYTHONPATH=... python -m pytest -q tests/frontend -> 28 passed
• earlier in this branch: PYTHONPATH=... python -m pytest -q tests/regression -> 4 passed

Notes

• row_prod / col_prod were removed from tile.__all__ because the installed PTO dialect does not provide backing builders for them; direct attribute access remains unchanged.