cathar v0.4.2

Backport of the LSAR de-clip improvement from 0.5.2 to the 0.4 line.

Changed

• De-clip now reconstructs clipped peaks with least-squares autoregressive
  interpolation (LSAR) — the classical audio-restoration method (Janssen,
  Veldhuis & Vries, 1986) — instead of a cubic fill. An AR model is fit to the
  reliable audio either side of each clipped run (two-sided autocorrelation →
  Levinson-Durbin) and the gap samples that minimise its prediction error are
  solved for (banded normal equations via Cholesky), so a clipped peak is rebuilt
  toward its true amplitude rather than flattened to the shoulder level. A
  stability guard falls back to the previous smooth fill when a solve rings or
  overshoots, so badly-clipped material softens gracefully.