Fix installation order when processing .install files

[dra27]

#4657 exposed no fewer than 3 bugs! 🐛

• OpamSystem.install doesn't the file permissions correctly if the file already exists (since open_out_gen only uses the mode if it actually creates the file). The first commit here fixes that and also the case of a symlink being overwritten which is handled by OpamSystem.copy_file but wasn't being handled by OpamSystem.install.
• opam-devel.install installs a file called opam twice to lib/opam-devel (fixed in Fix opam-devel.install #4664)
• The updates in Make install faster by not scanning the whole switch when no install field is present #4494 changed the order in which sections were processed - the second commit here reverses (excuse pun!) that slip

NB List.flatten blows the stack w.r.t. the number of lists (which is fundamentally small) and the length of the longest sub-list - this would already blow the stack in an earlier List.map call, so I convinced myself that a crazy opam .install file with gazillions of entries is already exploding.

[AltGr]

Seems good! See my comment on interrogations about special cases for #4494 though.

[AltGr]

This is the right behaviour, but now I am wondering what happens if a package overwrites a directory installed by another package with a file 🤔
It shouldn't be worse than just overriding files, and since this DirTrack is done outside of this, it should correctly detect the deleted files; but this has consequences with #4494, since that only handles lists of files expected to be installed: it should also track anything below in case it encounters a directory right ?

That seems pretty minor, though.

[dra27]

I'm not sure what happens at the higher level, but OpamSystem.install will fail if the destination is a directory?

[rjbou]

Several kind of files are written in the install process:

• the pkg.config file that is package specific by name
• the files in .install file that are written in prefix/pkg/xxx
• the misc files in .install that need confirmation from user to be installed, as they are absolute paths

Other install/copy are done via install field commands not opam instructions

[rjbou]

I'm not sure what happens at the higher level, but OpamSystem.install will fail if the destination is a directory?

indeed, it is the first check

let install ?(warning=default_install_warning) ?exec src dst =
  if Sys.is_directory src then
    internal_error "Cannot install %s: it is a directory." src;
  if (try Sys.is_directory dst with Sys_error _ -> false) then
    internal_error "Cannot install to %s: it is a directory." dst;

[kit-ty-kate]

My original PR in #4494 was significantly changed so none of this is my code (this also explains why i didn't see these issues sooner) and I can't really review it properly, but from afar it looks ok.