BinnedAggregate: add a passthrough flag

[cboulay]

What

Adds a passthrough flag to BinnedAggregateSettings, so this node can be left in a graph but switched off at runtime.

There was no way to do that before. Every other conditioning stage has one — CommonRereference has mode="passthrough", Butterworth has order=0, SamplingDelayAlignment has filter_len=0, Slicer has an empty selection — so a pipeline that wanted to compare with and without binning had to rewire itself instead.

A separate flag rather than a sentinel bin_duration, because the intended use is a runtime toggle: a consumer switching this off because the view zoomed to a range where binning would cost detail should not have to remember and restore the rate itself. A sentinel would have destroyed it.

How

__call__/__acall__ short-circuit before the state is ever consulted, matching EWMATransformer and AdaptiveStandardScalerTransformer:

def __call__(self, message: AxisArray) -> AxisArray:
    if self.settings.passthrough:
        self._request_reset()
        return message
    return super().__call__(message)

The _request_reset() is the load-bearing part. _state.carry holds the raw samples of an open partial bin; if it survived a passthrough gap, the first bin after resuming would mix samples from either side of it. Short-circuiting alone would cause exactly that, because _hash keeps its pre-gap value when nothing hashes it. _request_reset() sets _hash = -1 directly, so the next real message is guaranteed to take the reset path and start from a fresh schedule and an empty carry.

It lives in __call__ rather than only in update_settings so the guard also holds when settings are pushed by assignment (proc.settings = replace(...)), which is the path a runtime consumer is most likely to take. passthrough is in NONRESET_SETTINGS_FIELDS to record that update_settings need not queue a reset of its own.

Note for downstream

Toggling changes the shape of the output when operation is a tuple: the trailing metric axis appears and disappears with the flag. Anything downstream has to absorb a rank change — a fixed-layout sink reallocates, a plot rebuilds. ezmsg-tools' shmem bridge and sweep widget both do, but neither does it for free.

Deliberately not doing

EWMATransformer and AdaptiveStandardScalerTransformer resume from pre-gap state after passthrough, which is the opposite of what this PR does. That is intentional there and pinned by tests. The difference is that their state is a converged estimate, while the carry is raw samples that get spliced into an output bin. Whether theirs is right is #195.

Related: #196, an unrelated stale-cache bug found in RangedAggregateTransformer while comparing passthrough implementations.

Tests

Six tests in tests/unit/test_binned_aggregate.py:

• test_passthrough_forwards_untouched — output is the input object
• test_passthrough_keeps_the_bin_rate_for_when_it_is_switched_back — the flag-vs-sentinel argument
• test_passthrough_skips_the_state_entirely — no schedule built, no carry grown
• test_toggling_off_does_not_splice_stale_samples_into_the_first_bin — the one that fails without _request_reset()
• test_toggling_off_via_update_settings_also_resets — the other toggle path
• test_toggling_restores_the_metric_axis — the trailing axis comes back correctly

Full suite passes (3689 unit + 51 integration).