When 'src' needs to be moved to 'dst', but 'dst' already exists, the
command finds a suffix that doesn't exist yet.

This improves performance when dealing with large graphs.

As a result, fetchTaskFromQueueOrManager doesn't need the http.Request()
object pointer any more.

Works on files, directories, and non-existing paths.

The render output directory is moved to the same name, with its
modification timestamp appended in ISO-8601 format.

As a result, the render output directory MUST be known to Flamenco, and
thus from now on is a mandatory setting for the blender-render job type.

For now this is intended as a debug tool. It uses vis.js to visualise the
dependency graph of the job.

We need a format that can handle sub-second precision, which is not
provided by the HTTP date format (RFC 1123). This means that we can't use
the Last-Modified header, as it may be incorrectly interpreted and
rewritten by HaProxy, Apache or other software in the path between client
& server.

This format allows sub-second precision, which is what we need for reliable
queries.

When querying Server .../depsgraph, we just take the Last-Modified header
value as string, and store that without doing any parsing. This prevents
issues with timezones getting lost, formatting going wrong, etc.

Now the / url shows the version of Flamenco Manager that is running,
and 'flamenco-manager -version' shows that too (then stops).

When a manager gets /api/f/mng/{mng-id}/depsgraph, the returned jobs are
no longer automatically set to 'active'. This is delayed until a task
is updated to 'active', 'completed', or 'failed' (but only if that doesn't
fail the job yet).

If there is no task available for a worker, the worker now gets a quick
response that there is no task available. Simultaneously, the downloader
is kicked to fetch a new set of tasks from upstream Server.

Workers ask quite frequently for new tasks (every 5 seconds), so it is
likely that the downloader is already finished when a new request from this
worker comes in.

The new database name is immediately identifiable as ours, in contrast
to the generic "testdb".

The scheduling is actually performed by MongoDB using the aggregation
pipeline.

This commit also introduces QueueTaskUpdateWithExtra(), to push a task
update to Flamenco Server and set some extra fields in our local database
in one go.

Furthermore, in db.go I've added a type "M" as an alias for bson.M, so
queries look a little bit nicer.

Thanks to David Keeney for helping out with this.

Still rough & untested.

The /api/flamenco/managers/{manager-id}/depsgraph endpoint returns the list
of runnable tasks belonging to runnable jobs, assigned to that manager. All
tasks that were "queued" are set to "claimed-by-manager".