Update training init with finetune exports

Add finetune.py — finetune_domain_model (HF Trainer Pattern A, auto tabular_features passthrough)

Phase 2D: Fine-tuning pipeline — DomainFinetuneDataset, finetune_domain_model, 139 total tests passing

Add pretrain.py — pretrain_domain_model with HF Trainer, cosine schedule, DataCollatorForLanguageModeling

Add data_pipeline.py — tokenize_user_sequences, pack_sequences, prepare_clm_dataset

Phase 2C: Pre-training pipeline — data pipeline, sequence packing, HF Trainer CLM, 124 total tests passing