[Blobs V2] Bulk / coalesced reads for BlobFile

Motivation

Blobs are a crucial element for training workloads, where a dataloader requests a batch of 32-256 rows that include metadata as well as blobs (images, audio, etc.) per step and throughput matters more than random-access latency on any single blob. Currently, bulk-access performance is suboptimal: the only way to get concurrency from Python is wrapping individual BlobFile.readall() calls in a ThreadPoolExecutor.

Under the hood, take_blobs returns Vec<BlobFile>, where each BlobFile.read() issues an independent reader.get_range() call. This bypasses the FileScheduler coalescing logic that columnar reads benefit from. Reading N packed blobs that are sequential in the same .blob sidecar file results in N separate I/O operations, even when a single range read (or a few coalesced ones) would cover them all.

Proposed changes

1. Bulk read API

A method like Dataset::read_blobs(blob_files, concurrency) -> Vec<Bytes> that accepts the output of take_blobs and handles concurrent reads and coalescing on the Rust/tokio side, returning data directly. From Python this would be a single block_on call instead of requiring a ThreadPoolExecutor around individual BlobFile.readall() calls.

2. Read coalescing for blobs sharing a backing file

When multiple blobs reside in the same backing file (packed or inline), group them by path, merge adjacent/nearby byte ranges (similar to what FileScheduler::submit_request already does), issue fewer I/O operations, and slice the results back.

The Rust-side infrastructure for both improvements already exists (FileScheduler coalescing, futures::stream::buffered for concurrency) - it's mostly a matter of wiring it into the blob read path.

Related

#4947

Trace the existing take_blobs output through BlobFile.read/readall and the FileScheduler::submit_request path, then review futures::stream::buffered for the proposed concurrency flow. Use the requested Dataset::read_blobs API as the entry point; done means bulk reads return ordered bytes while blobs sharing a backing file use fewer coalesced range requests.