Handle Groq vision TPM limits

[arena-ai-coding-agent]

Summary

• cap vision completion tokens to fit Groq's 8K TPM on-demand tier
• retry token-budget 413 responses with progressively smaller output budgets
• compact schema prompts and document the vision token configuration
• preserve provider failures as retryable processing errors

Validation

• python -m py_compile backend/medical_extractor.py backend/api.py
• git diff --check