documents connector: add object-store (S3) input + predicate/prefix pushdown for scans

[abbccdda]

Context

The documents data source connector (crates/skardi/src/sources/providers/documents/) currently reads a local directory and re-parses the entire directory into memory on every SELECT (parsing is lazy at query time, with no caching, no streaming, and no pushdown). See the TODO(scale) / TODO(pushdown) comments in table.rs.

Two limitations to track here.

1. Add S3 / object-store input support

• Today the read path is local-filesystem only: collect_files() uses std::fs::read_dir and parse_file() passes a local path string to LiteParse::parse().
• The mod.rs doc comment already advertises "directory / object-store prefix" as the intent, and image_store (output side) accepts s3:// URIs but write_image_crop() is a no-op for remote stores — so the aspiration exists but nothing is implemented.
• Goal: let a source path be an object-store prefix (e.g. s3://bucket/prefix/) so cloud customers can point at their own bucket instead of having to land files on a mounted volume (PVC/hostPath) inside our deployment.
• Scope to consider: listing objects under a prefix, fetching bytes for parse, and finishing the deferred image_store upload path for remote stores.

2. Handle the full-directory scan case (predicate / prefix pushdown)

• TableProvider::scan() currently ignores its _filters: &[Expr] argument and does not implement supports_filters_pushdown, so DataFusion applies any WHERE above the scan as a separate FilterExec. A query like WHERE path LIKE 'batch-a/%' still parses the whole directory; even LIMIT 1 parses everything and then truncates.
• Goal: implement supports_filters_pushdown for path-prefix predicates and push the prefix into collect_files() (prune the walk, ideally only read_dir the matching subtree). Also consider limit pushdown and per-file streaming (RecordBatchStream yielding as each file parses) so large sources don't block a tokio worker while materializing one big batch.

Notes / references

• crates/skardi/src/sources/providers/documents/table.rs — scan() (ignored filters), fetch_partition() (eager parse, TODO(scale)/TODO(pushdown))
• crates/skardi/src/sources/providers/documents/parse.rs — collect_files(), parse_file(), write_image_crop() (remote no-op)
• crates/skardi/src/sources/providers/documents/mod.rs — "directory / object-store prefix" doc comment