Streamline settings management via ezmsg-baseproc delegation

[cboulay]

Summary

Adopts the new update_settings / NONRESET_SETTINGS_FIELDS machinery from ezmsg-baseproc#8 so this repo's units delegate live-update logic to their underlying processors instead of carrying it in unit-level on_settings overrides.

Net effect: CerePlexImpedance loses ~40 lines of bespoke override; live tweaks to freq_lo / freq_hi / test_current_nA no longer churn through a full processor recreate; the offsets-only fast path that preserves the accumulated impedance array is unchanged behaviorally but now expressed as a 10-line update_settings override on the processor.

Per-unit decisions

Unit	Change	NONRESET_SETTINGS_FIELDS
ChannelMapUnit	None — already inherits the right behavior	frozenset() (default; channel_map change should reset)
CerePlexImpedance	Drop unit on_settings override; add processor update_settings for the offsets-only fast path	{"freq_lo", "freq_hi", "test_current_nA", "headstage_channel_offsets"}
CereLinkSource	Untouched — async, multi-Session lifecycle doesn't fit BaseProducer.update_settings	{"cbtime", "microvolts"} (documentary; cont_buffer_dur deliberately excluded — consumed only inside open())

CerePlexImpedanceProcessor

• freq_lo, freq_hi, test_current_nA are read live in extract_impedance(...) and never cached in state → safe to update without resetting.
• headstage_channel_offsets requires re-laying out trackers but the accumulated state.impedance array remains valid; update_settings patches trackers in place via _build_trackers(n_ch).
• collect_duration_s and fft_duration_s are baked into state during _reset_state (buffer sizes, FFT window length); they correctly fall through the default reset path.
• The unit's prior on_settings override (and the dataclasses import) are removed; the inherited base-class default now routes through processor.update_settings(self.SETTINGS).

CereLinkProducer

NONRESET_SETTINGS_FIELDS added but documentary — the producer is non-stateful and CereLinkSource.on_settings overrides the base flow entirely (it manages two sessions across an async close → open, emits DeviceStatus, and rolls back SETTINGS on failure). The constant records which fields are safe to change live for future maintainers.

What's not changing

• CereLinkSource.on_settings — kept verbatim. The pycbsdk Session lifecycle needs await asyncio.to_thread, must avoid two simultaneous Sessions, and needs to publish DeviceStatus — none of which fits inside a sync update_settings(new_settings) -> None.
• ChannelMapProcessor — no code change. The new default behavior fixes a latent bug for free: today, swapping channel_map on a stream whose n_ch is unchanged would not re-parse the CMP because _hash_message only hashes n_ch. Post-PR, update_settings arms _hash = -1 so _reset_state re-runs on the next message.

Dependency

Requires ezmsg-baseproc with the new BaseProcessor.update_settings / BaseProducer.update_settings and the unit-level delegation in BaseProducerUnit.on_settings / BaseProcessorUnit.on_settings. The pinned range needs to be bumped to whatever version ships ezmsg-org/ezmsg-baseproc#8 once it merges (currently using a local editable checkout for development).

Test plan

• pytest tests/ --ignore=tests/test_impedance_integration.py --ignore=tests/test_integration.py — 44/44 pass
• ruff check src/ tests/ — clean
• New test test_offsets_only_update_preserves_impedance — verifies the
  accumulated impedance array survives an offsets-only update and trackers
  are re-laid out
• New test test_non_safe_field_arms_reset — verifies collect_duration_s
  change arms a full reset (_hash == -1)
• New test test_live_safe_field_does_not_reset — verifies
  freq_lo / freq_hi / test_current_nA updates don't reset state
• Integration test against live nPlay / hardware — settings push from GUI
  that toggles only microvolts should now apply without a Session restart
  (was previously silently ignored on idle, applied no-op on live session)