Development priority list

[jameskermode]

• Efficient jacobians (basis wrt positions) needed for ACE potential basis assembly for training
• Backward pass through full Lux-based ACE models
• Cleanup EquivariantTensors
• TRACE
• Bring back purification

Moved to ACEpotentials:

• Fully non-linear ACE fits - end-to-end Zygote-based AD through Lux-based ACE models
• Pre-trained/exported embeddings (e.g. from MACE) then used for linear ACE models or TRACE models
• Deprecate then drop support for pre-Lux models

[cortner]

• cleanup: remove unused kernels
• cleanup: rewrite kernels specialized to Rnl, Ylm to general embeddings
• refactor: remove A layer from sparse ACE basis; since A specifies the message, while the ACE basis is the update.
• discuss whether to just move everything to KA