Force single-threaded motion correction.

This is also to make the output deterministic, at the cost of running slow.

It turned out that using dense sampling wasn't enough; there was still some
numerical instability that came from the order of addition:

* https://github.com/ANTsX/ANTs/wiki/antsRegistration-reproducibility-issues#variance-due-to-floating-point-precision-errors

For some reason it only appeared on OS X, and only about 10% of the time, and never on Linux.
I [showed](#2642 (comment))
that the instability in isct_antsSliceRegularizedRegistration did exist on Linux, so something
still unknown about how we call it was hiding it there.

See:

* https://github.com/ANTsX/ANTs/wiki/antsRegistration-reproducibility-issues
* ANTsX/ANTs#444 (comment)
* ANTsX/ANTsR#210 (comment)

spinalcordtoolbox/moco.py

@@ -724,12 +724,9 @@ def register(param, file_src, file_dest, file_mat, file_out, im_mask=None):
         # run command
         if do_registration:
             kw.update(dict(is_sct_binary=True))
-            env = dict()
-            env.update(os.environ)
-            env = kw.get("env", env)
             # reducing the number of CPU used for moco (see issue #201)
-            env["ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS"] = "1"
-            status, output = sct.run(cmd, verbose=1 if param.verbose == 2 else 0, **kw)
+            env = {**os.environ, **{"ITK_GLOBAL_DEFAULT_NUMBER_OF_THREADS": "1"}}
+            status, output = sct.run(cmd, verbose=1 if param.verbose == 2 else 0, env=env, **kw)
 
     elif param.todo == 'apply':
         sct_apply_transfo.main(args=['-i', file_src,