fix(infer): add max_tokens guard + --trust-remote-code for SGLang

[kushdab]

Summary

Two targeted fixes for infer.py to address hanging inference and SGLang startup errors.

---

Fix 1: max_tokens guard in request payload (fixes #24)

Problem: When inference is run on certain images (sparse tables, high-contrast layouts), the model enters a detection-token repetition loop and never emits EOS. Because infer.py passes no token limit to the SGLang server, generation hangs indefinitely.

Fix: Add "max_tokens": CONTEXT_LENGTH to the request payload. This re-uses the existing CONTEXT_LENGTH constant (default 32768) so the guard is already configurable via --context_length. A natural EOS before the limit is completely unaffected; only runaway generation is stopped.

payload = {
    ...
    "max_tokens": CONTEXT_LENGTH,  # prevents infinite generation loops
}

---

Fix 2: --trust-remote-code in SGLang server launch (fixes #12, part of #27)

Problem: SGLang raises ValueError: Model architecture UnlimitedOCRForCausalLM is not supported on startup because the custom model class is not registered in the standard SGLang model registry.

Fix: Pass --trust-remote-code when launching the server, which tells SGLang to load the model class from the repo's modeling_unlimitedocr.py (same mechanism as Transformers).

python -m sglang.launch_server \
  --model baidu/Unlimited-OCR \
  --trust-remote-code \   # <-- added
  --port 30000

---

Testing

• Tested the hang fix by passing an image that previously caused the loop: generation now terminates at the token limit with a partial but useful output, and the next file proceeds normally.
• SGLang server starts without ValueError with --trust-remote-code.

Closes #24 · Part of #27 · Ref #12