Add configurable per-layer LoRA ranks (Related to Issue #2335)

[kor80]

Overview

This pull request introduces configurable static per-layer LoRA ranks, allowing users to assign different ranks to individual transformer blocks while preserving the existing global rank behaviour when no layer-specific configuration is provided.

Motivations

Static allocation makes it possible to assign different parameter budgets to individual transformer blocks without introducing additional training-time machinery. The allocation is explicitly defined before training, making configurations simple and reproducible.

Changes

• Add optional lora_r_by_layer configuration with input validation;
• Support block-specific ranks across attention, MLP and MoE modules;
• Retain the global rank for unspecified blocks and the language modeling head;
• Integrate the new option into the fine-tuning CLI;
• Support serialisation and restoration of layer-specific ranks through checkpoint metadata;
• Add test and usage documentation.

Testing

Added tests covering rank allocation, configuration validation, backward compatibility, checkpointing sanity check and training updates.

Limitations

• Rank allocation must be specified manually rather than learned or adapted during training.
• This implementation provides flexible parameter allocation but does not automatically identify which layers benefit the most from adaptation.

Closes #2335