Fix(parquet): Fix spurious refill failure for multi-page DATA_PAGE_V2 chunks

[r7raul1984]

Summary

PageReader::prepareDataPageV2's row == kRepDefOnly branch called
skipBytes() on a page's compressed bytes that had already been fully
consumed a few lines above via readBytes(). When a column chunk holds
more than one DATA_PAGE_V2 page and the first page's compressed size
exceeds the bytes remaining in the chunk after it, that redundant skip
drained the stream past the true end of the already-fetched buffer.
PageReader's logical byte offset (pageStart_) still believed more
data existed, so preloadRepDefs() issued one extra readPageHeader()
call, which failed with Thrift deserialize error: Empty buffer returned when refilling once the stream had nothing left to give.

This surfaced as a deterministic (non-flaky) failure when exploding a
nested list/array column whose chunk spans multiple data pages — e.g.
Spark's ParquetIOSuite test "explode nested lists crossing a rowgroup boundary" against packed-list-vectorized.parquet.

prepareDataPageV1 does not have this bug — it has no analogous extra
skip after consuming the page via readBytes().

The fix removes the redundant skipBytes() call; readBytes() already
advances past the whole page.

[r7raul1984]

@majetideepak hi, could u kindly take a look?