Fix merge: fall back to warm-start adapter from HF when GRPO skipped

Fix: batch_size=4 so num_generations=4 divides evenly

Fix: max_length -> max_seq_length for trl 0.15.2 (verified all configs locally)

Fix: add health server on port 7860 to prevent timeout

Fix: batch_size=16, 10K samples, unbuffered output, 2 epochs

Fix all: writable /tmp cache, no login(), proper permissions

Add Dockerfile and training script