EWMA/scaler: add opt-in reset_on_resume for passthrough gaps

[cboulay]

Addresses #195, as an opt-in setting rather than a behaviour change.

The problem

EWMATransformer and AdaptiveStandardScalerTransformer short-circuit passthrough in __call__/__acall__, returning before _hash_message, so self._hash keeps whatever value it had when passthrough was switched on. Switching passthrough back off therefore resumes from the state as it was before the gap.

zi is an exponentially-weighted average over the recent past. During passthrough the filter sees none of the samples that go by, so on resume it describes a window that ended when passthrough began — and the gap is invisible in the state, so a 10 ms blip and a 10 minute outage resume identically. For the scaler that means z-scoring post-gap data against pre-gap mean and variance.

The change

reset_on_resume: bool = False on EWMASettings, inherited by AdaptiveStandardScalerSettings.

Default False preserves current behaviour — the three tests pinning seamless resume are untouched, and the existing passthrough docstring promise still holds. A short blip arguably should resume seamlessly, and resetting discards an estimate that may have taken many time_constants to converge.

True rebuilds from the first post-gap message instead, matching the position BinnedAggregateTransformer already takes deliberately for the same flag. With the existing bias correction the first post-gap output is exactly the first sample, and the estimate re-converges over time_constant. For the scaler, _reset_state rebuilds both child EWMAs, so a parent reset propagates without extra work.

Two details

The reset is driven from __call__/__acall__, not update_settings. So it also holds when settings are pushed by assignment (proc.settings = replace(...)) rather than through update_settings. Both passthrough and reset_on_resume stay in NONRESET_SETTINGS_FIELDS: a toggle with no messages in between leaves no gap and so should not reset.

The passthrough and empty-message short-circuits are now separate. They were one or:

if self.settings.passthrough or np.prod(message.data.shape) == 0:

An empty chunk carries no samples past the filter and so leaves no hole in zi's history; only passthrough does. Sharing one condition would have made every empty chunk look like a gap. Both are now routed through a small _skip() helper that documents the distinction, with a test pinning that an empty chunk — including one arriving during passthrough — leaves _hash untouched.

Tests

New, all exercising reset_on_resume=True (the default path is covered by the existing tests):

• test_ewma_reset_on_resume_discards_state — output after the gap matches a never-used transformer, not one carrying the pre-gap estimate. Also asserts the state is left alone during the gap; only the hash is invalidated, so the rebuild happens on the next real message rather than eagerly.
• test_ewma_reset_on_resume_via_update_settings — toggling queues no reset by itself; the first passthrough message does.
• test_ewma_reset_on_resume_ignores_empty_messages — the case the split exists to protect.
• TestAdaptiveStandardScalerPassthrough::test_reset_on_resume_discards_state — both child EWMAs are replaced on resume.

Full suite: 3693 passed, 5 skipped. Ruff clean.