Update accumulated observable correction on the wide decode path

[ciaranra]

Found by an adversarial review of the PR #338 merge reconciliation: the wide-observables migration left StreamingLogicalDecoder's accumulated correction behind. The narrow flush() set accumulated_obs, but the wide flush_obs() did not — and the Python bindings route LogicalAlgorithmDecoder.decode() through the wide path, so accumulated_obs() returned 0 after a decode() that returned a nonzero correction. Only the narrow path had test coverage.

Fix (single decode path):

• accumulated_obs field is now a wide ObsMask, updated in flush_obs(); flush() delegates to flush_obs() and narrows with the same fail-loud error style as decode_to_observables (no truncation).
• accumulated_obs() -> Result<u64, _> (errors beyond 64 observables instead of truncating); new accumulated_obs_mask() returns the lossless mask.
• Python: accumulated_obs raises past 64; new accumulated_obs_mask returns an arbitrary-precision int via the existing obsmask_to_py helper.
• Regression test wide_decode_path_updates_accumulated_obs covers the actual bug (wide decode → accumulated state, plus reset), alongside the updated narrow-path tests.

Verified: 130 decoder-core tests pass (incl. the new regression), clippy cold-clean on both changed crates, fmt clean.