Why:

Building the benchmark object (setup_experiments in particular) can be
quite expensive (>60s). When we only want to execute one particular
analysis, the build does a lot of useless fetch on the storage. The
experiments should be built only when they are needed.
Relatedly, the analysis method should allow selecting specific study.

How:

Only call setup_experiments on study when we try to access
`study.get_experiments()`. The task_ids which seemed useless are removed
from `experiments_info` as well.

Why:

The interface is bloated and difficult to memorise. One of the reason we
decided to go this way is that we previously needed to instantiate an
algo to access its attributes. We can now use algo_factory.get_class()
to have the class object only and access the class attributes. Thanks to
this we can now simply set deterministic as a class attribute for the
algorithms and it simplifies the interface of the benchmark.

Algorithms have deterministic set to False by default since they
generally have sources of variations.

- Add gunicorn settings from YAML to Gunicon app
- increase gunicorn timeout to 300 seconds

… in with-statement, to make sure lock.acquire and lock.release are called

…ngoDB)