retrochimera 1.3.0

This release includes support for model finetuning, consensus mode for ensembling, stability improvements, and support for newer Python versions.

Changed

• Make template extraction more robust (#28) ([Krzysztof Maziarz (@kmaziarz)])
• Improve the data filtering script (#30) ([Krzysztof Maziarz (@kmaziarz)])
• Remove dependency on syntheseus-root-aligned and vendor relevant utils (#27) ([Lukas Geiger (@lgeiger)])
• Do not load template library in SMILES Transformer training (#29) ([Lukas Geiger (@lgeiger)])
• Update protobuf dependency to 5.29.6 (#25) ([Krzysztof Maziarz (@kmaziarz)])
• Use default factories for nested dataclass fields (#31) ([Lukas Geiger (@lgeiger)])
• Remove unused dependencies and unpin W&B (#37) ([Lukas Geiger (@lgeiger)])

Added

• Add support for model finetuning (#21, #24) ([Krzysztof Maziarz (@kmaziarz)])
• Implement consensus mode for ensembling (#26) ([Krzysztof Maziarz (@kmaziarz)])
• Allow ensembling to consume eval outputs produced in resumable mode (#32) ([Krzysztof Maziarz (@kmaziarz)])
• Expose the forward model class externally (#23) ([Krzysztof Maziarz (@kmaziarz)])
• Add config to set precision of torch fp32 matrix multiply (#36) ([Lukas Geiger (@lgeiger)])

Fixed

• Fix process leak in rule application server (#33) ([Krzysztof Maziarz (@kmaziarz)])