feat: cursor-based streaming for CSV export at scale (100K+ rows)

[kvithayathil]

Problem

_generate_csv_rows materializes all match results via list(yield_per(...)) before generating rows. This works at current scale (<10K) but will hit memory limits at 100K+ rows.

Approach

• Replace materialized list with cursor-based pagination through _generate_csv_rows
• Stream CSV rows via StreamingResponse with backpressure
• Yield rows in batches to avoid unbounded memory growth
• Preserve existing export format and behavior

Sub-tasks

• Design streaming API contract (chunk size, cursor state)
• Implement cursor-based iteration in _generate_csv_rows
• Wire StreamingResponse in campaign router export endpoint
• Integration test with 100K+ rows (in-memory SQLite benchmark)
• Verify memory stays bounded under load

Context

Identified as deferred follow-up from PR #124 (cursor-based keyset pagination). Low priority — current scale is <10K rows.