Make JavaReadable's coalescing gap configurable from Java #94

JavaReadable hard-codes its coalescing config to a 64 KiB gap and a 16 MiB maximum (vortex-jni/src/io/read_at.rs:92-98). Java readers, the Spark connector included, can't change it.

Why it matters. Files are laid out column by column, in runs of about 2 MB (BufferedStrategy in vortex-file/src/strategy.rs). So a scan that reads a few rows from many columns issues roughly one request per projected column buffer, and a 64 KiB gap rarely merges them. On object storage, each request is a GET.

Measured on real 476k-row files, reading the rows of one key (about 6 rows) with a 58-leaf projection, with the reads coalesced as JavaReadable does:

With many GETs and in-region reads, 1 MiB is far cheaper. About 14 requests is the floor at any gap.

Proposal.

• Take the gap, and optionally the maximum, from HadoopReadable / NativeReadable, and default to today's 64 KiB.
• Expose them as read options, e.g. vortex.read.coalesce.distance and vortex.read.coalesce.maxSize, which VortexIo passes through.
• The comment's concern stays valid: a wider gap can re-read segments the file's segment cache already holds. So the default stays as it is, and callers opt in.