v0.3.2: CMA-anchored learned evolution

This release adds CMA-anchored learned evolution with frozen meta-trained weights
and reproducible BBOB and Brax examples.

Added

• Added CMAAnchoredLES, a CMA-anchored learned evolution strategy with its
  frozen 364-parameter meta-trained checkpoint, packaged as portable NumPy pickle
  arrays with source provenance and exact weight-identity checks.
• Added a standalone, resumable eight-task Brax transfer evaluation example
  for CMAAnchoredLES, preserving the paper's legacy-spring protocol.
• Added CMAAnchoredLES to the BBOB notebook comparison and strategy table,
  with a Colab link.

Fixed

• Normalized source identifiers in Brax evaluation checkpoints across platforms,
  including Windows.