Add missing pyrefly suppressions to fix main CI

[kartik1pandey]

Fixes #1756.

Summary

pyrefly check fails on main with 10 errors across 4 lines. Each is a pre-existing type-stub precision gap (not a real bug — confirmed by running the affected test suites), missed when this codebase migrated from pytype to pyrefly (#1674) or after a subsequent JAX typing change (cf. 0f6d10b).

Changes

• linear_algebra.py: suppress missing-attribute on old_mat_h.dtype — identical pattern to the line immediately below it, which already has this suppression.
• linesearch.py: suppress unsupported-operation on decrease_error > atol — decrease_error is typed as ArrayLike (includes complex), but is only ever assigned real values.
• _make_pert.py: suppress bad-argument-type on baseline passed to vmap — baseline can be None, but is only read inside an if use_baseline: guard, so the None is never dereferenced.
• projections.py: suppress bad-argument-type on unravel_fn(new_values) — new_values is always a jax.Array at runtime; the union with ndarray comes from ArrayLike's breadth.

Verification

• pyrefly check --python-interpreter-path <venv>: 0 errors (was 10; 482 suppressed vs. 472 before).
• pytest optax/_src/linesearch_test.py optax/perturbations/_make_pert_test.py optax/projections/_projections_test.py -q: 295 passed, 1 skipped.
• optax/_src/linear_algebra_test.py::test_matrix_inverse_pth_root reproduced standalone (its module needs flax, unavailable in my sandbox): all condition-number cases pass with identical error values to the documented thresholds.
• Checked for conflicts with other open PRs touching these files (add missing docstring to get_spectral_radius_upper_bound in linear_algebra.py #1721, feat(perturbations): rename sigma→scale, default num_samples=1 (#1342) #1662) — neither overlaps the changed lines.

[kartik1pandey]

Closing this as superseded — the pyrefly failures this PR targets were independently fixed on main in d01841f ("Fix pyrefly type-check failures in the OSS CI"), which covers all 4 files touched here (and in two cases with a proper type fix rather than a suppression). No remaining diff to contribute.