CASSANDRA-19988 Honor background_read_disk_access_mode across SSTable maintenance and streaming

[cheeeee]

CASSANDRA-19988: extend direct-I/O reads from compaction to the other background operations that scan
whole SSTables, and rename the setting accordingly.

Changes:

• compaction_read_disk_access_mode is renamed background_read_disk_access_mode (Config @Replaces; the
  old key is still accepted and reported as deprecated). Java accessors are renamed with it; the
  ticket's setting shipped in 6.0-alpha only.
• Scrub, verify, SAI, SASI and Accord route index builds open their data readers with the configured
  background mode, as compaction already did.
• Partial streaming (CassandraStreamWriter, CassandraCompressedStreamWriter) reads stored bytes through
  StreamingFileReader, which under direct mode opens the file with ExtendedOpenOption.DIRECT and stages
  reads in a 1 MiB buffer aligned to the filesystem block size (block size is looked up once per
  directory and cached), handling unaligned section starts and the unaligned tail at end of file.
  Checksums and compressed chunk boundaries are served from the staged bytes exactly as before.
• Entire-SSTable streaming: sendfile from an O_DIRECT descriptor is either bounce-buffered or served
  from the page cache by the kernel, so under direct mode components are read through
  StreamingFileReader and written in 64 KiB batches; when the direct open did not actually take effect
  the reader is closed immediately and the zero-copy sendfile path is kept.
• Fallback is decided from the actual open: if the JDK cannot enable direct I/O on the file (it sets
  O_DIRECT with fcntl after opening and throws IOException when the kernel rejects it) or the block
  size is unusable (0, not a power of two, or larger than the staging buffer), StreamingFileReader
  falls back to a buffered channel and logs a rate-limited WARN naming the file and reason. The
  streaming path inspects no errno values and probes no temporary file; scrub, verify and index builds
  go through the pre-existing FileHandle.supportsDirectIO() check that compaction already used.
• cassandra.yaml/cassandra_latest.yaml document the new name, that on macOS DIRECT is only a cache
  retention hint (F_NOCACHE) and that tmpfs accepts it without effect; NEWS.txt gets a deprecation
  entry for the old key.

Testing:

• BackgroundStreamingTest (4): the raw reader returns identical bytes in standard and direct mode
  across unaligned start/end boundaries of a file in the SSTable's data directory; compressed and
  uncompressed partial streams reproduce the stored chunks/checksum slices; an entire-SSTable direct
  stream reproduces every component of a compressed and an uncompressed SSTable.
• CassandraEntireSSTableStreamWriterTest (2) and YamlConfigurationLoaderTest.backgroundReadModeAcceptsLegacyName.
• Kernel-level check, Linux 6.19: on ext4 an O_DIRECT read of 8 MiB leaves 0 pages resident and an
  unaligned read is rejected with EINVAL; on tmpfs O_DIRECT is accepted but all pages stay resident
  (which is why the yaml calls it a no-op there); on a filesystem without O_DIRECT support the kernel
  rejects the flag with EINVAL, which the JDK surfaces as the IOException the fallback catches.
• Not measured: streaming or scrub throughput under direct mode, and the page-cache effect on
  foreground reads. Incoming (receive-side) streaming is not covered by this change.

CASSANDRA-19988