L40S optimization: batch 8, disable gradient checkpointing, parallel dataloader

Speed optimization: 1 epoch, larger batch, shorter seq, no packing

Fix f-string format specifier error in logging callback

Reduce batch size to avoid OOM on L40S 48GB

Fix SFTTrainer: use tokenizer instead of processing_class

Use separate model name for cloud training

Use single GPU with larger batch size for L40S 48GB

Add 7B QLoRA training + AWQ quantization scripts