chore: reduce OpenAI vision costs for receipts

[marioommg]

Summary

• Compress/downscale receipt images in the browser before S3 upload to cut vision tokens and bandwidth.
• Send receipt images with detail: "low" and cap max_tokens on extraction responses.
• Switch title→category extraction from gpt-3.5-turbo to gpt-5-nano.
• Update README: document GPT-5 nano (no longer GPT-4 Vision) and the cost-saving behavior.

Test plan

• Enable NEXT_PUBLIC_ENABLE_RECEIPT_EXTRACT and upload a receipt photo; confirm amount/date/title/category still extract correctly.
• Upload a large photo (>2MB / high res) and confirm it is compressed before upload (smaller file / JPEG).
• Enable NEXT_PUBLIC_ENABLE_CATEGORY_EXTRACT and verify category suggestion from title still works.
• Confirm README describes GPT-5 nano and cost notes accurately.