Native Multi-Channel Support for classify_segments and Streaming Detection

Second standard building block toward real robotics adoption. Every real robotics experiment in this repo -- LeRobot's 6 joints (Experiments 43, 46, 47), the UCI HAR IMU's multiple axes (Experiment 41) -- has needed a hand-written for j in range(n_channels): loop around a function built for one 1D signal. Real robots always have more than one channel. This is ergonomics, not a new algorithm: classify_segments_multichannel and MultiChannelStreamingDeviationDetector apply the SAME, already-validated per-channel logic across all channels at once.

Correctness bar: identical to the hand-written loop, not "close enough"

et_mc, dev_mc, unc_mc = classify_segments_multichannel(X, classify_segments, **ARBITER_KW)
# vs. calling classify_segments(X[:, j], **ARBITER_KW) for each j and stacking by hand

Verified on real, multi-channel data from two independent domains -- the same bar velocity_gated_stable_mask and Experiment 48's streaming detector were promoted at:

LeRobot episode 0 (all 6 joints) (batch, n=303, c=6): all_channels_match=True
LeRobot episode 0 (all 6 joints) (streaming, n=303, c=6): all_channels_match=True
LeRobot episode 22 (all 6 joints) (batch, n=237, c=6): all_channels_match=True
LeRobot episode 22 (all 6 joints) (streaming, n=237, c=6): all_channels_match=True
IMU (real UCI HAR, 3 axes, n=3072) batch match: True
IMU streaming match: True

Zero mismatches, both the batch wrapper (classify_segments_multichannel) and the streaming one (MultiChannelStreamingDeviationDetector, built on Experiment 48's StreamingDeviationDetector), across a real 6-joint robot arm and a real 3-axis IMU.

Design note: channels are independent by construction

Each channel gets its own reference window and its own baseline -- a robot's joints or an IMU's axes are not assumed to share one statistical baseline (a fast-moving joint and a near-stationary one have very different natural noise scales; forcing a shared baseline would either desensitize the quiet channel or false-alarm on the active one). This is a design choice stated explicitly, not an oversight: cross-channel relationships (does channel A's anomaly correlate with channel B's) are a different question, tested and rejected as a detection mechanism in Experiment 46 -- this module does not attempt that, it only removes the need to hand-loop over independent per-channel detection.

Reproducing this

scripts/dense_armor_streaming/multichannel.py (classify_segments_multichannel, MultiChannelStreamingDeviationDetector) and scripts/dense_armor_streaming/validate_multichannel.py -- reuses already-cached LeRobot and UCI HAR data, no new downloads.