fix(export): add time/size guards to dumpDatabaseRoute for large DBs

[CuboYe]

/claim #59

Problem

/export/dump loads the entire database into memory before returning a file. For large databases (>1 GB, approaching Durable Object limits), this causes OOM or hangs past the 30s request timeout with no error message or partial result.

Fix

src/export/dump.ts — three layers of defence:

1. Pre-export size estimate — runs pragma_page_count + schema size query before loading any row data. Fails fast with 413 if estimated dump >100 MB, with a clear message pointing to per-table fallbacks.

2. Per-table runtime budget (10s) — checked between rows inside the INSERT loop for each table. Catches tables with millions of rows mid-export, before they consume all memory.

3. Global time budget (20s) — checked between tables. Stops export cleanly after N tables and tells the caller exactly which table caused the overflow and how many remain.

Error messages are actionable: they name the failing table, the row count if hit, and direct users to /export/csv/:tableName / /export/json/:tableName as a workaround.

What didn't change

• The core dump format and INSERT generation logic are untouched — only guards added
• Existing tests still pass (the small-DB path is identical behavior)
• No new dependencies or R2 bindings required for the in-memory path

Testing

• Verified pragma_page_count query works in the existing test harness
• Time budget paths tested by injecting Date.now() mock via Vitest
• Large-DB error path (413) tested with estimatedBytes threshold mock

[CuboYe]

Checking in — any questions on the approach? Happy to adjust the time budget thresholds or add R2 streaming if preferred.

[CuboYe]

Friendly bump — PR is ready for review.