v0.0.54: VLM ORPO OOM fix, ZeRO-3 device_map fix, env var rename

Add VLM support for ORPO trainer (v0.0.53)

Subclass ORPOTrainer as VLMORPOTrainer to support vision-language
models (e.g. Qwen 3.5-9B) with image+text preference data.
Ports DataCollatorForVisionPreference from DPO to ORPO, adds
image_column param, and wires up AutoProcessor loading.

Bump version to 0.0.52

Fix ORPO/DPO crash on pre-formatted data with empty input_ids (v0.0.51)

Skip input_ids pre-tokenization for ORPO/DPO trainers since they use
chosen/rejected columns, not text. Empty input_ids from missing text
column crashes transformers floating_point_ops() with numel error.

Fix Qwen3.5 tool_calls crash, token leak defense, dep upgrades (v0.0.50)

- Auto-parse JSON string arguments to dicts in safe_apply_chat_template
  for models like Qwen3.5 that use arguments|items Jinja filter
- Add __repr__ masking for token/wandb_token in AutoTrainParams
- Scrub HF tokens from log files before upload_folder
- Upgrade deps: trl>=0.29.0, accelerate>=1.13.0, peft>=0.18.1,
  huggingface_hub>=1.6.0, sentence-transformers>=5.2.3, transformers>=5.3.0

Fix Qwen3.5 support and 409 repo conflict handling (v0.0.49)

- Fix _check_tool_calls_support for Qwen3.5 (try both string and dict)
- Add exist_ok=True to push_to_hub create_repo calls
- UploadLogs: detect 409 conflict, create datetime-versioned repo,
  propagate versioned repo_id to config for final push
- Sync chat templates from unsloth 2026.2.1 (42 templates)
- Add actual Qwen3.5 template from tokenizer (43 total)
- Add sync_chat_templates.py script
- 8 tests in test_repo_conflict.py

Fix _check_tool_calls_support for Qwen3.5 (v0.0.48)

- Qwen3.5 template uses arguments|items which requires dict not string,
  try both in probe to avoid false negative
- Add exist_ok=True to all HF Hub create_repo calls (fix 409)
- Sync chat templates from unsloth 2026.2.1 (42 templates)
- Add actual Qwen3.5 template extracted from tokenizer (43 total)
- Add sync_chat_templates.py script
- Add transformers-dev optional dependency

Fix ORPOConfig import for TRL 0.29 compat (v0.0.47)

ORPOConfig was removed from trl top-level in 0.29, moved to
trl.experimental.orpo. Use try/except fallback to support both
0.28 and 0.29+.

Remove max_prompt_length from ORPO/DPO configs (v0.0.46)

TRL 0.28.0 moved ORPOConfig to trl.experimental and removed
max_prompt_length. DPO deprecated it too. Prompt length is now
inferred from max_length - max_completion_length.

Fix ORPO/DPO multi-turn prompt extraction (v0.0.45)

Use explicit prompt column when available instead of always deriving
from chosen[:-1], which breaks multi-turn where the completion spans
multiple turns.