Improve performance with copy_file_range(2)

On file systems supporting reflinks or extent remapping, such
as XFS, Btrfs, NFS, regions of files that are not modified are
processed without any copying — with O(1) cost.

While at it:

- Correct the use of write(2) to handle partial writes and
  report appropriate errors.

- Improve buffer sizing.

These are the benchmark results
           no-op-short copy-file-range is  1.17 times faster than main
      access-log-no-op copy-file-range is  2.91 times faster than main
   access-log-no-subst copy-file-range is  1.56 times faster than main
      access-log-subst copy-file-range is  1.03 times faster than main
     access-log-no-del copy-file-range is  1.59 times faster than main
    access-log-all-del copy-file-range is  1.06 times faster than main
   access-log-translit copy-file-range is  1.05 times faster than main
access-log-complex-sub copy-file-range is  1.04 times faster than main
     access-log-append copy-file-range is  1.05 times faster than main
             remove-cr copy-file-range is  1.03 times faster than main
          genome-subst copy-file-range is  1.08 times faster than main
            number-fix copy-file-range is  1.03 times faster than main
           long-script         main     is similarly fast as   copy-file-range
                 hanoi         main is  1.03 times faster than copy-file-range
             factorial copy-file-range is  1.03 times faster than main

On suported file systems (XFS, BTRFS), output files with no or few in-place
modifications share large blocks with the input file, as can be seen
in the following example.

$ xfs_io -c 'fiemap -v' lines.txt
lines.txt:
 EXT: FILE-OFFSET      BLOCK-RANGE         TOTAL FLAGS
   0: [0..857199]:     11917304..12774503 857200 0x2000
   1: [857200..857207]: 12774504..12774511      8   0x1

Author: dspinellis