Stream video uploads to disk instead of buffering whole files

[luojiyin1987]

Summary

The video upload endpoint currently calls await file.read() and only checks the size after the entire upload has been buffered in memory. With a 500MB limit, one request can consume hundreds of MB in the app process before rejection.

Impact

• High peak memory usage per upload request
• Concurrent uploads can exhaust FastAPI worker memory
• Oversized uploads are rejected too late, after the buffering cost has already been paid

Proposed fix

• Write uploads to disk in chunks
• Count bytes while streaming and reject as soon as the request exceeds MAX_FILE_SIZE
• Clean up the temp session directory on size-limit or write failures
• Add regression coverage for successful streamed uploads and oversize rejection

Notes

A fix branch already exists in my fork: luojiyin1987:fix/stream-upload.