This aligns the behavior with the opam loader.

@aleo it's actually a good idea that I did this slight adaptations ;). Now I recall that
for the deployment [1], it would work but often passing through the opamroot
initialization step in the loader. Because, there is no guarantee that the actual worker
listing the package (with a given opamroot as /tmp/<something>) will be the same one
ingesting the package. Thus, the opamroot passed along to the loader may not exist, so
it would get initialized and possibly multiple times.

We should probably make that opamroot a configuration entry in the dedicated
configuration file targeting a fixed location in the workers to avoid the too many
opamroot locations in /tmp (for another time/diff).

[1] T3438

Related to T3358