fix: defer heavy imports in inference CLI for vLLM compat

[slacki-ai]

Summary

• Move torch, vllm, transformers, huggingface_hub, and openweights.client imports from module top-level into the __main__ guard
• This allows monkey-patches to be applied before vLLM is imported — vLLM captures tqdm and tokenizer behaviour at import time, so patching after import has no effect
• Adds tqdm noise reduction (rate-limited updates) and transformers.PreTrainedTokenizerBase.all_special_tokens_extended compat patch for newer transformers versions
• Changes main() signature from main(config_json: str) to main(cfg, conversations) — config parsing and data loading now happen in __main__ before vLLM import

Changes

• openweights/jobs/inference/cli.py — restructured import order and main() signature

Test plan

• AST-based unit tests verify import structure (11 tests): heavy imports not at top level, stdlib imports preserved, main() signature correct, __main__ guard contains deferred imports
• Integration: run an inference job end-to-end to verify the new import order works with vLLM

🤖 Generated with Claude Code

[nielsrolf]

Why do we need this? I strongly prefer imports at module level. Modifying tqdm behavior is imo probably not worth such a change. Why does the tokenizer need any monkeypatching?