Keep the input verbatim in mri/orig/001.<ext> , and write mri/rawavg.mgz for the tools

[m-reuter]

mri/orig/001.mgz was written by run_prediction.py through nibabel, re-encoding the input
into MGH. For a scaled NIfTI that write raised inside a discarded Future, so the file was
silently absent; otherwise it was a lossy record of what had been processed. mri/rawavg.mgz
did not exist, so recon-surf.sh linked it to the conformed orig.mgz and pctsurfcon
measured the gray/white contrast on the conformed image instead of on the input.

Two files now, with distinct jobs:

• mri/orig/001.<ext> is a byte-for-byte copy of the input, under the extension it arrived
  with, placed by run_fastsurfer.sh. Nothing reads it back. Formats spread over several
  files, such as an Analyze .hdr and .img pair, are copied whole. A T2 is copied the same
  way as mri/orig/T2raw.<ext>, the name recon-all uses.
• mri/rawavg.mgz is what the tools read. It is a relative symlink to the archival copy when
  the input is already an .mgz, and a conversion through save_image otherwise, which sets
  fov. The T2 equivalent is mri/orig/T2raw.mgz, which N4 now reads.

Also in this PR:

• The copy runs for both pipelines, so a seg-only run leaves behind the rawavg a later
  surf-only run needs, and before the LIT module, so the contrast is measured on real tissue
  rather than on inpainted voxels.
• An archive of a different image stops the run. The comparison is byte for byte, so a
  different image under the same extension is caught too. One subject directory belongs to one
  input.
• --base writes neither file, since its input is the template the pipeline built and it skips
  pctsurfcon. --long writes rawavg only.
• long_prepare_template.sh imports each time point through copy_input.py rather than
  mri_convert, so the only copy the longitudinal stream keeps of a time point input is
  verbatim. Both consumers read the resulting rawavg.mgz.
• run_prediction.py holds the conformed-image write in a deque the main loop drains, so a
  failure in the pool reaches the exit code.
• SubjectDirectory.copy_orig_name and its plumbing are removed.

Known consequences:

• ?h.w-g.pct.stats is now measured on the input. With a zero minimum, conform's rescale is a
  pure multiplication that the contrast cancels, leaving quantisation noise.
• Base directories no longer contain mri/rawavg.mgz.
• Re-running into a directory that already holds an archive of a different image fails instead
  of overwriting. The message names the archived copy to pass in order to reprocess the same
  input, and asks for a subject id of its own for a different image.

Tests: test/image/test_copy_input.py covers the verbatim copy, both rawavg paths, multi-file
inputs, the two conflict cases and --rawavg_only. expected-files.yaml gained a glob section
for outputs whose name depends on the input.