refactor: centralize async runtime around shared loop and pools

[hrodmn]

The concurrency model was somewhat messy (as noted in #31) so I wanted to make things simpler. Here are a few refactor-ish changes that make lazycogs concurrency easier to reason about:

• DuckDB chunk queries now yield the event loop by running through one bounded internal helper and explicit executor.
• The sync bridge now submits to one shared persistent background loop instead of per-thread loops, and first-use loop startup waits until that loop is actually running.
• Reprojection now uses one explicit bounded pool, and the store cache no longer depends on thread-local state.

The single reprojection pool affected the benchmark conditions for the single-worker case so we see a regression in that benchmark test.

I also removed the not-very-useful read_chunk synchronous shim around read_chunk_async.