[format] Fix RLE skipping at sparse row-range boundaries

[CrownChu]

What is the purpose of the change?

An RLE run covers the half-open interval [rowId, rowId + n). When its end is exactly equal to rangeStart, it has no overlap with the requested sparse row range and must be skipped. The existing strict < boundary can retain that run and misalign sparse Parquet reads.

For example, consider an RLE run starting at row 0 with n = 10 and a sparse read requesting row range [10, 10]:

RLE run:     [0, 10)   // rows 0 through 9
Sparse range: [10, 10] // row 10

These ranges do not overlap. However, the old condition rowId + n < rangeStart evaluates 10 < 10 to false and enters the overlap branch. It then computes start = 10, end = 9, and an invalid zero-length intersection. Using <= skips the preceding run and lets decoding continue at row 10.

Brief change log

• Change all three affected RLE boundary checks from < to <=.
• Add a data-evolution regression test covering sparse row positions.
• Verify both a flat column and a repeated/list column at the boundary positions.

The regression writes 10,000 Parquet rows with an INT column and an ARRAY<INT> column, then performs a sparse indexed read at row positions 10 and 4210. It verifies the exact output:

1000:[0,1]
5200:[4200,4201]

This exercises both the flat definition-level path and the repeated/list path across Parquet page and RLE boundaries.

Verifying this change

mvn -pl paimon-core -am -DskipITs -Dfast \
  -Dtest=DataEvolutionSplitReadTest \
  -Dsurefire.failIfNoSpecifiedTests=false test

Result: 14 tests passed. Spotless and Checkstyle also passed.

[leaves12138]

+1