Engine reload floor: cache readable batches, build tables from columns

[nullarch]

Reloading a database after new events land costs about 0.45 s even with the incremental path in place. The expensive part is no longer decoding (segments are cached, and only newly listed ones are decoded) but rebuilding the query engine:

• projection tables ≈ 292 ms
• readable event batches ≈ 124 ms

Both are rebuilt from scratch on every refresh, and both are cacheable.

What to do

In crates/attemptdb-query (cache.rs, tables.rs):

1. Cache readable Arrow batches per segment the way ScanCache already caches decoded segments, so a refresh concatenates cached batches instead of re-encoding every event.
2. Build the projection tables from typed column builders over the projection that the incremental projector already produced, instead of going through the generic TableBuilder row-by-row path.

Done when

cargo run -p attemptdb-bench -- step refresh shows a materially lower reload time at 200 k events (today: 0.5 s against 5.6 s from scratch), the numbers land in docs/benchmarks.md, and cargo test -p attemptdb-query -p attemptdb-ui -p attemptdb-mcp -p attemptdb-server stays green — the cache is shared by all four.

Performance work with a benchmark that already exists to prove it. Not a first issue, but well bounded.