Reuse manifests within each data source

[platypii]

Repeated queries on one Iceberg data source currently fetch and decode the same data manifests on every scan. Retain each manifest's promise within that source's lifetime, preserving lazy partition pruning and sharing in-flight reads across concurrent scans. Failed reads are removed so later scans can retry; separate sources retain independent snapshot state.

Add tests for lazy loading, repeated and concurrent scans, source isolation after appends, and retries. Also cover equality deletes across data/delete manifest merges, schema renames, manifest rewrites, and reads of the original snapshot.

Validation:

• All 943 tests pass; lint and TypeScript checks pass.
• On a local HypAware table, repeated queries now read 26 manifests on the first scan and zero on subsequent scans.
• Release review compared eight local HypAware tables covering approximately 1.2 million rows and 128 query comparisons across unchanged, rewritten, and merged manifests, with no result differences. Original tables were not modified.

Loaded manifest entries remain in memory while their data source is reachable; there is no shared cache across source instances.