fix(jni): cache each open file's segments and merge reads across small gaps only - #66

Summary

Closes #65. Spark reads of list-layout files read several times the bytes they need, because the JNI data
source opens files without a segment cache and a list column's element segments are requested once per row range
of a scan. On a 47 MB list-layout file of nested event data, a 29-column read now reads 13.7 MB instead of 39.5 MB,
and 15.7 MB instead of 68.8 MB when Spark splits the file in four. Before this change, the same read of the file
written without the list layout read 30.0 MB, so the list layout now pays off instead of costing more.

Changes

• Each file the JNI data source opens gets its own segment cache, capped at 64 MiB and dropped with the open file.
  Both NativeDataSource.open and openFiles use it.
• JavaReadable merges reads across gaps of at most 64 KiB instead of 1 MiB. With a cache, a merged read that
  spans a segment the cache already holds reads it again; at 1 MiB that left list-layout reads at 1.3x to 1.8x
  the bytes they need, at 64 KiB 1.0x to 1.2x.

Testing

• writer::tests::list_layout_scan_reads_each_segment_once: writes a list-of-struct column with the list layout,
  scans it as four row ranges through the data source, and checks that no segment is read twice with the JNI open
  options. It also checks that without them some segments are, so the test keeps reproducing the problem.
• cargo test -p vortex-jni, cargo clippy -p vortex-jni --all-targets.
• Spark, local mode, the connector on develop with this native library, same file, bytes the JVM read for the
  29-column read: list layout 21.5 to 25.0 MB before, 14.5 MB after; split in four, 77 to 92 MB before, 48.4 MB
  after. Without the list layout: 30.6 MB before, 29.5 MB after. Row counts match.

Not fixed here

Spark reads each split of a file in its own task, so splits can't share a cache, and an elements segment that
spans the whole file is read once per split. Reading a list-layout file in one split avoids that
(vortex.splitFiles=false); writing list elements in chunks that line up with the outer row chunks would fix it
for split reads too.