tp: keep a pipeline's rows when SQLite reads it again

[LalitMaganti]

SQLite reads the inner side of a join once for every row of the outer side. When a pipeline is on the inner side it runs again each time, including every pipeline it reads through SQL, which makes a join against a pipeline quadratic.

This adds Memoize, a source which passes on another source's batches while keeping them, and replays them on later reads once a read has reached the end. A read stopped part way, or one whose batches can't be kept, keeps nothing.

The pipeline table function puts one in front of a plan the second time the same cursor reads it, so a pipeline read once pays nothing extra and one read any number of times runs at most twice.

[github-actions]

🎨 Perfetto UI Builds & Tests

• db3cedc (2026-09-27 06:50 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36300964322-1-ui/ui/index.html
• 05a69cc (2026-09-27 16:47 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36333280544-1-ui/ui/index.html
• 7dde121 (2026-09-27 18:16 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36337696599-1-ui/ui/index.html
• 2ea9583 (2026-09-27 19:28 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36343809583-1-ui/ui/index.html
• 9b79f01 (2026-09-27 20:43 UTC): https://storage.googleapis.com/perfetto-ci-artifacts/gh-36345310866-1-ui/ui/index.html

[LalitMaganti]

Closing for now. Inputs are now collected once, so this caching may not be worth it any more; if it is, it will come back on top of #7644.