Related T703

This resolves the caveat mentioned in prior commit about not being
able to use blake2 prior to 3.5

Related T692
Closes D192

- Add module docstring
- Add blake2s256 and blake2b512 in supported algorithms
- Spawn a new variable DEFAULT_ALGORITHMS as default computed
  algorithms for the main functions

Related T692

Remove the limit on the python3 version, this should be transparent.
If the hash requested is not available, this will raise with an
explanation on the error.

Related T692

The same caveat applies, will only be supported from python3.6 onward.

Related T692

The caveat is that it will only be supported when we will be using
python3 >= 3.5.

Related T692

Related T700

Closes T255
Ref. D179

Found by investigating T75

The subversion loader (T680) has shown that throwing floating point values
around for timestamps is a mess waiting to happen.

We now coerce all clients to send us timestamps as integer numbers of seconds
and microseconds, avoiding data losses everywhere.

- default
- ignoring empty folders
- ignore folder based on pattern in names

This permits to reuse the same logic for different clients (loader-dir,
loader-tar, loader-svn)

(Tests were lost)

This actually is supposed to replace
walk_and_compute_sha1_from_directory.
The data structure used here is better at handling updates.

(Code that actually got lost and rewritten - Tests are definitely lost though)

Keep the old api (since i don't measure the impacts on other modules yet).

+ Improve docstring

(Code that actually got lost and rewritten)

Combination of:
- validation on files
- ignore empty folder

In some corner case, the changed paths can reference a previous ignored
folder (thus not existing in the data structure)

The empty folder was not previously in the objects structure.
So we need to add it as child of its parent for the update.

Enforce convention on directory name without trailing /.
At the moment, the `git.walk_and_compute_sha1_from_directory` injected
the rootdir with a possible trailing / (input from client).

Since we compute it anyway, better return it along with the result

Decrease the number of paths to compute to 1 common ancestor (if any):
- Scan only that directory and rehash with new results (data changed)
- Update the resulting objects with those new hashes.
- Update from that directory to the rootdir the existing hashes computation