env-var step overrides for ad-hoc test runs

fix: lora_dropout=0.05 + HF gradient checkpoint to bypass unsloth GRPO dtype bug; chdir to /tmp before merged push

resilient training: SFT push insurance, dtype workaround, try/except GRPO, SKIP_GRPO env

fix: writable OUT_DIR (/tmp), chat-template text column, push_history_to_hub

fix: render chat template into text column for new unsloth/TRL SFTTrainer

fix: add unsloth_zoo to bootstrap step

deploy v5 (long mode, a100x4)

deploy v5 (long mode, a100x4)

deploy v5 (long mode, a100x4)

deploy v5 (long mode, a100x4)

deploy v5 (long mode, a100x4)

deploy v5 (long mode, a100-x4)

deploy v5 (long mode, a100-large)

initial commit