[tx] Simplify LoRAMixin for linear and embedding case

[pcmoritz]

We simplify the LoRAMixin by disentangling the linear and embedding case and handling them in LoRAMixin and LoRAEmbed respectively. This is in preparation for supporting LoRA for the flattened layout (tokens, *dims) in addition to (batch, seq_len, *dims) and also #969

[gemini-code-assist]

Code Review

This pull request refactors the LoRAMixin to better separate the logic for linear and embedding layers. The introduction of the _compute_lora_intermediate method, which is overridden in LoRAEmbed, is a good use of polymorphism and simplifies the apply_lora method by removing conditional logic. My review includes a couple of minor suggestions to improve code clarity by marking unused function parameters, which is a good practice for maintainability.