[M4] NumPy zero-copy and type stubs

[tetsuh]

Summary

Implement zero-copy NumPy reads (get_array) and ship type stubs.

References

• docs/05_api_python.md §1 (type mapping), §2.2 (get_array), §4 (stubs)
• docs/01_requirements.md P02

Scope

• python/bindings/numpy.cpp: ParamCache.get_array(key, dtype=...)
  returning a read-only ndarray over the cached buffer
  (buffer protocol; writeable=False; keepalive holds the
  shared_ptr<const ParamValue>)
• Size/dtype mismatch raises TypeMismatchError
• ndarray put: C-contiguous arrays accepted via buffer protocol
  (one copy at encode time)
• python/sitos/py.typed + .pyi stubs for the whole public API;
  mypy runs in CI
• tests/python/test_numpy_zero_copy.py: base-pointer identity
  (no copy), writeable flag, keepalive under value overwrite
  (Python-side SpanHandleSurvivesOverwrite equivalent)

Out of scope

• dtype/shape metadata in the payload (not part of wire v1)

Acceptance Criteria

1. get_array returns a view (no copy — verified via base pointer) with
   writeable == False
2. Array stays valid after the key is overwritten (keepalive)
3. mypy is green against the stubs

Dependencies

• [M4] Python ParamCache #24

For the implementer

• Copy the checklist above into your PR description and tick items as you
  complete them. Do not edit the checklist in this issue.
• Follow docs/development_workflow.md (branch issue/27-numpy-zero-copy,
  Conventional Commits per §2.1, TDD RED first).
• This issue is closed automatically by the PR via Closes #27.