Draft: Global Salt refactoring

This MR breaks backward compatibility with pre-GN3 models (like GN1 and GN2*)

Summary

This MR introduces a global refactoring of salt/models. It aims to remove old unsupported code and transfer all legacy models to the same transformer backend, as used in GN3.

Models refactoring

• Remove GATv2 attention completely
• Replace old transformer and attention modules with their analogs from transformer_v2
  • muP options are supported
  • edges are supported via separate EdgeAttention class
• No CrossAttentionPooling, wasn't used anywhere

Backward compatibility policy

• In all configs replace TransformerEncoder with Transformer.
• Match num_layers and num_heads
• Attention -- flash-varlen
• Norm -- pre with LayerNorm
• Dense -- gated=False
• Dropout for attention -- not supported

Conformity

• Documentation Updated?
• Development Guidelines followed?
• Pipeline Passes?