v0.3.18: bug fix with colqwen2

[0.3.18] - 2026-08-22

Added

• Add the optional [plaid] extra (fast-plaid>=1.4.7,<2.0.0 and fastkmeans) for faster matching on larger corpora, and include it in the [all] extra.

Changed

• Extend supported dependency ranges to allow torch<2.14.0, peft<0.21.0, and pillow<12.4.0 in the training extra.

Fixed

• Fix ColQwen2/ColQwen2_5 checkpoint conversion to remap model.embed_tokens and model.norm to the language_model.* layout used by Transformers v5, so these weights are no longer silently dropped and randomly re-initialized when loading from Qwen2-VL checkpoints.