feat: asynchronous scan prefetch (Phase 3a MVP) - #17

Closes the Phase 3a Rust test gaps:

Prefetcher-unit (compute_ranges):
- split-range pruning matches the opener's first-page rule
- statistics pruning drops a row group with filterAware=true, keeps both
  with filterAware=false
- nested struct projection maps to the right leaf column chunks
- encrypted spec degrades to footer-only
- CSV spec covers exactly the split (bounded / whole-file / empty)
- footer-parse failure returns Err and run() counts prefetch_files_skipped

Integration (ProbeStore counting/delaying/failing inner store):
- prefetch drives concurrent upstream requests (max in-flight > 1 — the
  apache#3817 single-outstanding-I/O fix)
- a second scan of the warmed ranges issues zero upstream requests
- an injected upstream failure is counted (prefetch_errors) while the
  demand read still succeeds (errors are never cached)

Refactors the existing warming test onto shared helpers (write_parquet /
store_over / parquet_spec / lcg). All native gates stay green.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Adds PrefetchFromS3Suite (extends the existing MinIO testcontainers
harness) exercising the full JVM -> JNI -> native prefetch path over a
remote (s3a) store — the scheme the data cache actually wraps:

- prefetch ON: a native Comet scan warms the cache ahead of the decoder;
  getDataCacheStats shows bytes_fetched > 0 and blocks_consumed > 0
  (observed 15 MiB fetched, all 15 blocks consumed, 0 wasted/errors),
  results byte-identical to the expected aggregate.
- prefetch OFF (probe): the same cold scan still fetches through the
  cache but issues zero prefetch requests / bytes — proving the prefetch
  counters are attributable to the prefetcher, not the cache.

Also folds in a spotless reformat of the CometConf prefetch entries.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

The demand path (`fill_pass`) fetched coalesced runs in a sequential
loop, so with the data cache in the path a scan's independent
column-chunk fetches were serialized — a regression versus vanilla
`object_store::get_ranges`, which coalesces and fetches a request's
ranges in parallel. Drive the runs through `buffer_unordered`
(cap DEMAND_FETCH_CONCURRENCY) and publish each as it lands, so waiters
still wake promptly and the mutations stay single-threaded.

Independent of prefetch: this restores parallelism the cache dropped.
With it, the scan prefetcher's contribution is purely fetch<->decode
overlap rather than also compensating for a serial demand path.

Co-Authored-By: Claude Opus 4.8 <[email protected]>

Adds a cache+prefetch cold-scan column beside the existing cold/warm
arms (§5/§6). Per scenario it now reports Cold (cache-only), Cold+PF
(cache+prefetch, same ColdFetch bytes), PF-Speedup, Warm, and prefetch
Coverage (consumed / (consumed+wasted)).

Because the demand path now parallelizes coalesced fetches, both cold
arms parallelize, so Cold vs Cold+PF isolates the fetch<->decode overlap
specifically — the number that classifies a workload: Cold+PF ≈ Cold =>
bandwidth-bound (no overlap to reclaim); large PF-Speedup =>
fetch<->decode-serialized (decode/latency-bound). Prefetch is toggled
per-query via the SQLConf (read per-plan). Runs against the same
env-configured s3a base path as the rest of the benchmark.

Co-Authored-By: Claude Opus 4.8 <[email protected]>