OpenEphysBinaryRawIO: detect timestamp gaps and add gap_tolerance_ms

[h-mayorquin]

OpenEphysBinaryRawIO previously had no gap detection (there was a literal # TODO for later : gap checking), so a within-recording discontinuity, for example from a dropped USB packet, was silently treated as contiguous and every sample after the gap got the wrong time. This adds gap detection using the per-sample sample indices the format already stores in sample_numbers.npy (v0.6+) or timestamps.npy (pre-v0.6): any non-unit increment is a gap. Detection is exact, since these are integer sample counters with no jitter.

The interface matches the cross-reader gap API from #1773 that is already in Blackrock and Neuralynx. gap_tolerance_ms=None (the default) raises a ValueError with a per-gap report, so the reader never hands back a silently mis-timed recording; a float segments the recording at gaps larger than the threshold and absorbs smaller ones (0.0 segments on every gap); and ignore_integrity_checks=True bypasses the check for fast loads or corrupt-file recovery. It also adds _get_openephysbinary_timestamps for direct access to the raw per-sample indices, matching _get_blackrock_timestamps and _get_neuralynx_timestamps.

One intended behavior change: loading a recording that contains a gap now errors by default instead of returning a single silently-concatenated segment. Clean recordings are unaffected; code that was unknowingly reading a gapped recording will need to pass gap_tolerance_ms or ignore_integrity_checks=True. The tests synthesize gaps from a real clean fixture (openephysbinary/v0.6.x_neuropixels_with_sync), since no public OpenEphys recording with a real gap exists.

@alejoe91