v0.41.0

Fixed

• Q2_K dequantization now matches ggml — fixes corrupt Q2_K output: both
  Q2_K dequant impls used a "16 sub-blocks reading qs[j*4]" scheme that applied
  the wrong super-block scale to the wrong 2-bit lanes, producing corrupt F32
  output (185/256 elements wrong vs ggml/candle on a representative block).
  This silently corrupted every Q2_K/Q2_K_S model (common on HF) — both the
  format path (apr tensors/inspect/validate/convert) and the inference
  path (apr run/serve). Fixed both (aprender-core format dequant +
  aprender-serve inference dequant) to match ggml dequantize_row_q2_K /
  candle BlockQ2K::to_float byte-for-byte (golden falsifiers FT-Q2K-001/002,
  contract contracts/q2k-dequant-parity-v1.yaml).