[ktf]

Once a grid file was open, the native reader gave up on the first failed read,
and it sent a whole read plan at once, with plain reads of up to 64 MiB. On
Hyperloop train 767444 at ORNL, which reads most files remotely, native
subjobs failed on I/O errors from Poznan EOS: in 8 failure logs examined, 6
were "Operation expired" (a request taking longer than XRD_REQUESTTIMEOUT,
150 s in the ALICE environment) and 2 "bad address". ROOT does not retry
these either; it only moves to another replica when opening fails.

• FailoverByteSource reads the replicas of a file: a read failing with an
  IoError, at any point of the file, is retried on the same replica, reopened,
  up to Options::readRetries (2) times per replica with a doubling delay from
  Options::retryDelay (2 s), and then continues on the next replica of the
  same size. Format errors are not retried. openNativeAOD() now goes through
  ByteSource::openReplicas() instead of trying the replicas only at open.
• XRootDByteSource asks the data server for its kXR_readv limits, as
  TNetXNGFile does, splits large ranges into plain reads of at most 8 MiB, and
  sends a plan in rounds of at most Options::remoteInFlightBytes (32 MiB), so
  that no request has to wait for a whole plan on a slow link.

Tested: the AODNative unit tests, with new ones for FailoverByteSource (retry,
failover, size mismatch, open failures, concurrent readers), and the
AnalysisSupport and DataInputDirector tests. o2-test-framework-aod-native-vs-root
on alien:///alice/sim/2025/LHC25b4a6/2/559362/AOD/003/AO2D.root (2 DFs) from a
Mac: the CERN EOS replica does not exist on the server, both readers moved to
the next replica, and the native result is byte-identical with ROOT's. Not yet
measured on the Grid.

[ktf]

Tested on hyperloop. No regressions.