fix: avoid scheduler deadlock on same-priority chunks

[Xuanwo]

This fixes a scanner hang when a ranged read starts at a write-batch boundary and a wide column is split into multiple I/O chunks while another smaller, higher-priority request is scheduled in the same scan line.

The scheduler previously used only the minimum in-flight priority to decide whether a request could bypass byte backpressure. If a higher-priority request remained unconsumed while the decoder awaited a wide-column request, remaining chunks from the already-admitted wide-column request could be blocked behind the byte budget, causing a deadlock.

This changes both the standard and lite schedulers to let chunks with an already in-flight priority keep moving. It also adds scheduler-level and scanner-level regression coverage, and updates the io_buffer_size docs now that this deadlock is fixed.

Validated with the relevant lance-io scheduler tests and the targeted scanner regression.

[codecov]

Codecov Report

❌ Patch coverage is 81.70732% with 15 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-io/src/scheduler.rs	77.61%	15 Missing ⚠️

📢 Thoughts on this report? Let us know!

[westonpace]

Hmm, too late to chime in but I wonder if there is a way we could have made all parts of the logical request have the same priority? Where was this split happening?