Stabilize Sybil doctests against JAX scalar repr drift

[Copilot]

The failing unxt Python 3.13 on macos-latest job was caused by brittle doctest expectations, not a quantity-operation regression. JAX scalar Array(...) reprs vary by environment (weak_type=True, optional trailing dtype metadata), and Sybil was treating those formatting differences as failures.

• Root cause

  • Doctests in Python docstrings and docs assumed a single exact scalar JAX repr.
  • On the failing runner, valid outputs differed only in optional repr metadata.

• Doctest normalization

  • Added a custom Sybil/doctest output checker in conftest.py.
  • Normalizes JAX scalar repr noise before comparison:
    • strips optional weak_type=True
    • normalizes doctest placeholders like dtype=float32, ... so they still match when no extra metadata is present

• Scope

  • Keeps runtime quantity behavior unchanged.
  • Applies only to doctest evaluation, so existing examples remain readable while becoming portable across JAX/Python/platform combinations.

• Coverage

  • Added focused unit tests for the checker:
    • accepts repr-only variation
    • rejects substantive output changes

Example of the normalized mismatch this addresses:

# expected in doctest
Quantity(Array(2., dtype=float32), unit='m')

# actual on some runners
Quantity(Array(2., dtype=float32, weak_type=True), unit='m')