fix: bound multi-image chat payloads

[adity982]

Summary

• keep the original uploaded image unchanged in workspace storage
• generate a bounded vision preview (512 KiB, 1568 px maximum edge) for the chat payload
• normalize EXIF orientation and transparent images while compressing oversized inputs
• cover unchanged small images and bounded large-image previews with regression tests

Why

The upload endpoint embedded each original image as base64 in the WebSocket message. A single allowed 10 MiB image expands beyond 13 MiB as base64, so a multi-image message can exceed practical frame limits and lose image context. Bounding only the model-facing preview keeps multi-image payloads manageable without degrading the stored original.

Validation

• python -m pytest tests/test_chat_upload_images.py -q — 2 passed
• uvx ruff check app/services/chat_image_preview.py tests/test_chat_upload_images.py
• uvx ruff check --select I app/api/upload.py

Addresses #776.