Replace busy-wait publishers in ResampleConcat and ResampleUnit with event-driven publishing

[cboulay]

Summary

ResampleConcat.output() and ResampleUnit.gen_resampled() both polled next(self.processor) in a while True loop with await asyncio.sleep(0) when no output was ready. sleep(0) yields to the event loop but never sleeps, so each unit spun a full CPU core whenever idle. This PR makes both units publish only when output can actually exist.

ResampleConcat: drain from the subscriber handlers

The polling publisher is removed entirely. on_reference / on_signal each push to the processor and then drain it, publishing every chunk it can currently produce (same idiom as Align). This is lossless because the composed resampler is always reference-driven (resample_rate=None), and in that mode output readiness only ever changes on new input — the wall-clock max_chunk_delay extrapolation applies to prescribed-rate mode only. A single next() consumes all currently-eligible reference values, so the drain loop runs at most twice per message.

Side benefit: backpressure is preserved — a subscriber does not complete until its outputs are published, so a slow consumer backs up the input queues instead of silently growing the resampler's internal buffers.

ResampleUnit: event wake with a timeout in prescribed-rate mode

This unit cannot publish from its handlers alone: in prescribed-rate mode with a finite max_chunk_delay, output is intended to become ready by wall clock with no input. The publisher therefore waits on an asyncio.Event set by both input handlers, using max_chunk_delay as a wait_for timeout when resample_rate is set and the delay is finite. The event is cleared before draining so a push landing mid-drain re-arms the next wait rather than being lost. on_signal is overridden via @ez.subscriber(BaseConsumerUnit.INPUT_SIGNAL) (the same pattern Downsample uses) to add the wake.

Pre-existing issue surfaced (not fixed here)

While testing the timed wake I found that the max_chunk_delay wall-clock extrapolation is currently unreachable at the processor level, and was under the old polling design too: after a drain the source buffer retains only ~2 samples, and ResampleProcessor.__next__ returns early on src.available() < 3 before evaluating b_project. The timed wake preserves the intended trigger for when that guard is reworked (linear interp1d only needs 2 points, so relaxing the guard for the extrapolation path looks feasible); the new integration test documents this and can be extended to assert extrapolated output once fixed.

Tests

• tests/integration/ezmsg/test_resampleconcat_system.py (new): drives ResampleConcat through a live ez.run graph with interleaved reference/signal chunks at 100 vs 99.7 Hz; asserts output flows, has A+B channels, and that the published time axis stays strictly monotonic with both handlers publishing to the same stream. Terminates on quiet, which also proves nothing depends on a background publisher task.
• tests/integration/ezmsg/test_resample_system.py (new): drives ResampleUnit in reference-driven and prescribed-rate modes; asserts output flows and is monotonic, with a documented pointer for the extrapolation assertion above.
• Full suite: 3651 passed, 7 skipped.