Repositories stored in ~/.opam/repo contain thousand of tiny files
and directories, and can take a huge amount of time to load on HDDs,
networked or old filesystems. This is not visible in normal use
because there is a marshalled cache, but can cause opam update to be
extremely slow (it now needs to compute a diff on the files, which
requires re-loading the repository drom disk), or big lags when
changing your opam version (and the full file tree is not yet in the
OS memory cache).

The solution proposed here is extremely pragmatic, yet quite
efficient: we store the repository contents as .tar.gz files in
~/.opam/repo instead. Rather than resorting to a complex in-memory
structure, we just untar them to /tmp when they need to be read, and
re-tar them after modification (opam update, or format upgrade
only). Then we let the OS disk cache do the job: in normal operation,
the tree never needs to be flushed to disk, and loading the .tar.gz
is orders of magnitude faster than loading the individual files.

Note that this is done even for rsync or git repositories, which
is not particularly clean, but works. In the case of git, it would
be possible to just store a bare repository, and use git to extract
the individual files (this could even be done explicitely, directly to
memory, see how Camelus performs). But it does not seem worth the
additional implementation cost at the moment.