Add Parquet footer metadata cache for metrics scans

[alexanderbianchi]

Summary

• Wire metrics ParquetSource through DataFusion's CachedParquetFileReaderFactory so Parquet footer and page-index metadata can be reused across scans.
• Increase the shared Quickwit DataFusion runtime file metadata cache budget to 4 GiB per node.
• Add a session-builder override for the metadata cache limit so we can raise or tune it independently later.
• Add tests covering both the cached reader factory attachment and the 4 GiB runtime cache default.

Expected impact

This targets the current metrics IO bottleneck where repeated scans spend a large amount of time in Parquet metadata loading/opening before returning rows. Warm runs on the same worker should reduce metadata_load_time, time_elapsed_opening, and time_elapsed_scanning_until_data for repeated access to the same files.

This does not cache column data or object-store byte ranges. bytes_scanned should remain roughly unchanged until we add a full blob/range cache.

Follow-up

For a full file/range cache, DataFusion/object_store OSS gives us range APIs and range coalescing, but not a packaged production byte-cache. The likely Quickwit integration point is either the QuickwitObjectStore adapter or the existing quickwit-storage StorageWithCache/StorageCache path, backed by Foyer or a similar bounded cache.

Validation

• cargo fmt -p quickwit-df-core
• cargo test -p quickwit-df-core
• cargo test -p quickwit-datafusion