[Config] Add merge LoRA iteratively feature

[1danchirila]

Create a feature that allows LoRA adapters merging sequentially. The flow would be: Train on dataset_k with LoRA => merge_lora => Train merged model on dataset_k+1.

[flaviusburca]

Added a feature to load a LoRA adapter and merge it online with the base model weights. The option is 'adapter_path'. Not supported with pre-quantized models like MXFP4, NVFP4, FP8 as it would require special kernels.

888101e

[1danchirila]

Not implemented the way it was intended.