Aggregate: ARGMIN/ARGMAX return coordinates; fix RangedAggregate async passthrough

[cboulay]

Closes #192.

AggregateTransformer ARG* returns coordinates

AggregateTransformer returned ARGMIN/ARGMAX as a raw index into the aggregated axis, while RangedAggregateTransformer and BinnedAggregateTransformer return the axis coordinate for the same enum member. It is worse here than for the other two, because this transformer removes the axis from the output — the consumer is handed an index into a dimension the message no longer describes, and cannot convert it without having kept the input around.

The machinery was already in place; AggregateTransformer was opting out of it. This drops index_to_coordinate=False and widens the coordinate condition from op == TRAPEZOID to the existing needs_coordinates(op), which already covers both ARG* members.

d = np.zeros((1, 10)); d[0, 7] = 1.0     # peak at index 7 -> 14.0 Hz
msg = AxisArray(data=d, dims=["ch", "freq"],
                axes={"freq": AxisArray.LinearAxis(gain=2.0, offset=0.0, unit="Hz")}, key="k")

AggregateTransformer(AggregateSettings(axis="freq", operation=A.ARGMAX))(msg).data
# before: array([7])     <- index, and "freq" is gone from dims
# after:  array([14.])   <- Hz

Breaking. ARG* output units change from index to axis coordinate. TRAPEZOID was already correct and is unaffected, and no in-repo caller uses ARG* through either transformer (fbcca.py calls numpy.argmin directly). Where the axis carries no metadata, get_axis supplies a default LinearAxis (gain=1, offset=0), so the coordinates are 0, 1, 2, … and the result is the index after all — covered by a test.

index_to_coordinate stays on aggregate_slices for a caller that means to index back into the array it passed in; nothing in this package sets it False any more, and the docstring now says so instead of naming AggregateTransformer.

Separate bug: RangedAggregate(bands=None) was broken in every graph

Found while reading. RangedAggregateTransformer overrode __call__ for its bands is None passthrough but not __acall__. BaseTransformerUnit.on_signal awaits __acall__, so in a graph the shortcut was skipped entirely and bands=None fell through to _reset_state:

TypeError: 'NoneType' object is not iterable    # for start, stop in self.settings.bands

The shortcut only ever worked when the transformer was driven synchronously, which is why every existing test passed. BinnedAggregateTransformer overrides both.

Both branches now also call _request_reset(), matching BinnedAggregate, so toggling bands to None and back to a different list cannot resume on the first list's slices.

Tests

• test_aggregate_transformer_argminmax updated to expect coordinates (freq is gain=2.0/offset=1.0, so index and coordinate genuinely differ).
• test_aggregate_argminmax_matches_ranged — new, pins the actual invariant: a band spanning the whole axis asks the same question as the full-axis reduction, so the two transformers must agree. Guards against a future re-split.
• test_aggregate_argminmax_bare_axis_is_index — new, the no-metadata fallback.
• test_ranged_aggregate_passthrough_async / ..._resume_async — new, cover the __acall__ path.

Full suite: 3699 passed, 5 skipped. Ruff clean.