This also means that we need a `tests/__init__.py` file; this file makes
py.test undertand what is our sources directory, and thus allows them to
`import flamenco_worker`. As a result, the test imports from tests/*.py
need to change to relative imports.

Early detection of missing dependencies trumps a few ms of startup time.

This will help debugging the crazy double-Blender-running we're seeing
in the studio.

This prevents multiple subprocess commands running at once. They shouldn't
be doing that in the first place, but we have a strange bug where multiple
Blenders seem to be running on the same machine.

This check is performed every time before trying to fetch a task. It checks
for readability or writability of certain filesystem paths, to prevent
tasks from failing when the shared storage hasn't been mounted. Instead,
the worker will go to `error` status and sleep for 10 minutes before trying
again.

This makes the naming more consistent, and makes it easier to add more
functionality without growing an already-big function.

Cycles and other render engines produce a line of log for each object in
the scene. In the Spring project there are so many objects (leaves, twigs,
pebbles) that it chokes the Manager with logs.

For now we have some custom code to swallow those lines, in lieu of a
logging system that can handle those volumes properly.

Task updates would be sent in an infinite loop when the Manager didn't
know the task, blocking all other task updates.

It was quite useful in finding the recent memory leak.

I'm guessing the do_db_push() async function execution was starved by
other asyncio tasks, causing Python to remember all the to-be-queued
payloads in memory. By making the function synchronous this doesn't happen.

Not all systems have it.

We now collect 1000 lines of log instead of 10, and push activity and logs
every 15 resp. 30 seconds (instead of 1 resp. 5)

This should prevent an accumulation of task updates, when the updates
are generated faster than they can be sent to the Manager.

Also added a warning in case the queue size actually grew while flushing
to Manager.

This reverts commit 30094fce.