Community fork with relaxed dependencies and expanded test coverage

[volgin]

Hi! We've been using Boltz in production and hit dependency conflicts with the strict == pins, so we created a community-maintained fork at volgin/boltz-community.

What we've done so far:

• Relaxed dependency pins from == to >= for ecosystem compatibility
• Removed fairscale (not imported anywhere) and the numpy<2.0 cap
• Cherry-picked open community PRs: Fixed --write_full_pae and --write_full_pde options. Fixes #585 #602, Fix #583 parsing .csv MSA files removes paired sequences. #584, Fix #581 constructing MSA feature array #582, Removing content-type from authentication headers #488, fix: corrected sign error and clarified calculation of binding energy #363
• Added 108 new tests (CPU-only, no checkpoint required) covering loss functions, model layers, data parsers, potentials, and output writers
• Found and fixed a bug in PairformerLayer where the v2=False code path crashes at runtime due to unconditionally passing k_in= to v1 AttentionPairBias

We're also reviewing other open PRs (e.g. Apple Silicon support) for inclusion — some may need modifications, so we'll work with the original authors where possible.

PRs welcome — happy to upstream any of this if it's useful. The fork is MIT-licensed, same as the original.