CASSANDRA-21085 Preserve offsets when splitting cache eviction requests larger than 2 GiB

[cheeeee]

CASSANDRA-21085 reports "Failed trySkipCache on file: ... Error: Invalid argument" during compaction.

NativeLibrary.trySkipCache(fd, offset, len, path) splits a request into int-sized chunks for the native
binding. After the first chunk it subtracted the chunk length from the offset instead of adding it, so
every following chunk was issued with a negative offset and posix_fadvise returned EINVAL. This can only
happen when the requested range exceeds Integer.MAX_VALUE bytes, i.e. on data files larger than 2 GiB.
Separately, a zero-length request (meaning "to end of file") dropped its starting offset and was issued
from 0.

Changes:

• advance the offset after each chunk;
• keep the requested starting offset when len == 0;
• size chunks at 2 GiB rounded down to a 2 MiB multiple (FADVISE_MAX_CHUNK) instead of Integer.MAX_VALUE.
  POSIX_FADV_DONTNEED only drops pages fully covered by the range, so an unaligned chunk boundary leaves
  the page that straddles it resident; a boundary that is a multiple of every common page size
  (4K, 16K, 64K) does not.

Testing:

• NativeLibraryTest: testSkipCacheLargeRange (Byteman on the native call; asserts the exact (offset, len)
  tuples for a range of 2 * FADVISE_MAX_CHUNK + 17 bytes, three calls), testSkipCacheZeroLength (offset
  preserved), testSkipCache.
• Kernel-level check, Linux 6.19, ext4, 4 KiB pages, sparse 2 GiB + 8 MiB file: DONTNEED issued in
  Integer.MAX_VALUE chunks leaves the page straddling the first chunk boundary resident (mincore: 1 page);
  issued in FADVISE_MAX_CHUNK chunks it is evicted (0 pages).

CASSANDRA-21085