BinnedAggregate: allow a tuple of operations, stacked on a metric axis

[cboulay]

Lets BinnedAggregateSettings.operation be a tuple, applying several aggregations to the same bins and stacking the results on a trailing axis.

First of a stack of two. #194 builds on this branch.

Why

Putting a high-rate signal on a screen wants each bin's min and max. Stride decimation lands between samples and clips the peaks, so a spike shrinks or vanishes depending on where the stride falls; the two extremes preserve peak amplitude exactly. That was going to be a dedicated min/max decimator until it became clear the only thing it would do that BinnedAggregate does not is apply two functions instead of one — everything else (binning on the shared BinSchedule, carrying the open partial bin across message boundaries, labelling the output axis) already lives here and is better tested than a new transformer would have been.

What

BinnedAggregate(BinnedAggregateSettings(
    bin_duration=0.001,
    operation=(AggregationFunction.MIN, AggregationFunction.MAX),
))
# (time, ch) @ 30 kHz  ->  (time, ch, metric) @ 1 kHz

The trailing axis is named by newaxis (default "metric") and coordinate-labelled with the AggregationFunction values — "min", "max" — so consumers read names rather than relying on position.

Design notes

• Trailing, not in place, so the binned axis keeps its position.
• Output shape follows the type of operation, not the count. A scalar behaves exactly as before — no trailing axis, no new key in axes — so existing streams are untouched. A one-element tuple does produce the axis, so a caller assembling its tuple programmatically gets a stable shape either way.
• Computing them together is not just less code: the bins are cut once and sliced once, so the results are identical by construction rather than by two transformers agreeing on a grid. A test compares a tuple against separate single-op transformers over the same chunking to hold that.
• The empty path needed the axis at full width too. At a high input rate most chunks close no bin, so _empty_like is the common path, and a zero-length message of the wrong rank will not concatenate with the messages around it.
• The metric axis depends only on settings, so it is built in _reset_state and held in state; every output carries the same object, which keeps a downstream identity check cheap.

Tests

10 new, on top of the existing 20. 98 pass across test_binned_aggregate, test_aggregate, test_bin_schedule, plus test_empty_gate (the only other BinnedAggregate consumer).