Structured-pruning-adapters

Fixed
- Grouped convolutions

Added
- Low-memory `SPLoRALinear` forward.
- Structured Pruning Parallel Residual Adapter (SPPaRA)

Fixed
- `SPLoRAMultiheadAttentionPruner` usage by circumventing concat in SPLoRAMultiheadAttention impl (weird issue...)

Added
- `SPLoRAMultiheadAttentionPruner` for torch-pruning compatibility

Changed
- Moved named_parameters and parameters from class methods to independent functions

Added
- `torch-pruning` pruners for SPLoRA
- `SPLoRAMultiheadAttention`

Changed
- SPLoRA and SPLoPA behavior to automatically set weight.requires_grad = False instead of assertion
- Change SPLoPA weight initialisation to set LowRankMatrix near-zero instead of pos_weigts