added in arrays to collect per-atom descriptors and gradients

[jmgoff]

• Arrays B_all and neighbours_dB were added to collect descriptors and descriptor gradients, respectively.
• To break out descriptor contributions per atom, other intermediate arrays were added (weights_rank1_dB and weights_dB for gradient breakout).
• Descriptors and descriptor gradients may be collected and printed using the LAMMPS compute. This compute is implemented in the 'compute-pace' branch here.

[yury-lysogorskiy]

@jmgoff I added more general and controllable implementation of gradients for both ACEBEvaluator and ACECTildeEvaluator in feature/b-grad branch. Could you, please, check if that will work with your compute pace ?