Detect dropouts below 1e-15 V instead of requiring an exact zero - #218

Stacked on #217 (contains its commits; review the last commit, or merge #217 first and this will show only its own change).

Why

find_bad_by_correlation counts a window as a dropout only if the MAD of the filtered signal is exactly 0. Any filtering before or inside NoisyChannels leaves residue in a flat stretch:

• the default do_detrend=True high-pass alone: ~1e-21 V; only ~9 of ~30 dropout windows stayed exactly zero in a synthetic test,
• a typical 1 Hz high-pass + 100 Hz low-pass + notch chain before NoisyChannels: up to ~3e-15 V, 0 windows exactly zero, so bad_by_dropout never triggers.

The existing test_bad_by_dropout passes only because it uses do_detrend=False.

What

• A window is flat for a channel if the MAD of its filtered signal is below 1e-15 V, the same value as the default flat_threshold of find_bad_by_nan_flat. Signal MAD in EEG is ~1e-6 V, so the margin is about nine orders of magnitude.
• matlab_strict=True keeps the exact-zero criterion.
• Documented as a deliberate difference in the MATLAB differences page and in the method docstring; changelog entry.

Measured on synthetic data (30 s dropout, 1 Hz HP + 100 Hz LP + 50/100/150 Hz notch, DC offsets 0–300 mV): 16–20 of 28 dropout windows flagged (windows at the dropout edges contain filter ringing and are not flagged), i.e. far above the default frac_bad=0.01.

Tests

• A dropout is detected with the default do_detrend=True (fails on main).
• A stretch of 1e-17 V noise counts as a dropout by default (fails on main) but not with matlab_strict=True.

Full suite: 71 passed (including the MATLAB comparison tests); pre-commit clean; docs build clean with -W.