jni: split oversized Java reads into chunks instead of capping coalesce max size #104

Problem

JavaReadable::read_at rejects any read over i32::MAX bytes ("exceeds ByteBuffer limit"), because each read is one readFully upcall into a single Java ByteBuffer.

To keep merged reads under that limit, checked_coalesce_config (vortex-jni/src/io/read_at.rs) caps coalesceMaxSize at i32::MAX - (Alignment::MAX_UNTRUSTED - 1). This has two drawbacks:

• Coupled to another crate. The headroom copies how FileSegmentSource (vortex-file/src/segments/source.rs) pads the coalescing window. If that padding changes, the cap is silently wrong.
• Only covers merged reads. A single segment over 2 GiB still fails.

Proposal

Split oversized reads in JavaReadable::read_at into readFully upcalls of at most i32::MAX bytes each, every one over a slice of the same aligned buffer. JavaWritable already does this for writes (vortex-jni/src/io/write.rs, bytes.chunks(i32::MAX as usize)).

Then:

• Remove the MAX_UNTRUSTED clamp and its i32::MAX test case.
• Remove the "native code caps values near Integer.MAX_VALUE" caveat from the NativeReadable.coalesceMaxSize() Javadoc and java/vortex-spark/README.md.

Follow-up from #95.