Fix: export OOM on large crawls — add streaming export endpoint

[liquidpurple]

The current export flow has a 4× memory multiplication that causes OOM kills
on memory-constrained instances:

1. Frontend fetches ALL URLs from /api/crawl_status (full data in browser)
2. Frontend sends ALL data back in POST body to /api/export_data
3. Backend generates full export string in memory
4. Backend wraps in jsonify() JSON envelope
5. Frontend parses JSON, extracts content, creates Blob

For a crawl of ~2200 URLs (~45MB in SQLite), this consistently OOMs on a
2GB RAM instance with systemd MemoryMax.

This PR adds a streaming GET endpoint (/api/export_stream) that uses Python
generators to yield CSV/JSON/XML rows one at a time via Flask Response.
The browser downloads the file directly — no JSON envelope, no Blob,
no round-trip of all data.

The old /api/export_data endpoint is preserved for backward compatibility.

Changes:

• main.py: new /api/export_stream endpoint + 8 streaming generator functions
• web/static/js/app.js: exportData() rewritten to use streaming endpoint