This commit reworks the agent task actions so that they are compatible
with the configuration cache. The problem was that when the configuration
cache is enabled, it will try to serialize the task actions that we add,
so that if the task is not up-to-date and that the configuration didn't
change, it can replay the action.

The problem was that our task action used, as an input, a `Provider<List<String>>`,
where those strings represented paths which do _not_ exist when the
action is serialized yet. In practice, we don't care, because those
are implementation details of the task action: the directories will
not exist _before_ the task is executed, and not _after_, they only
exist during task execution.

The configuration cache, however, wants to know about all inputs.
We could have changed the action so that it uses a _root directory_
instead of a list of directories as input, and make the computation
of the `session-` directories _during_ the task action: this would
have worked because the only input would be the "output directory"
that we pass as root. However, the task action is also used in another
task which uses completely different inputs, which are not derived
from a single directory.

At the same time, we _don't_ need the `Provider` semantics in the
task action: all we care about is to have a supplier which will
generate a list of directories. Therefore, we changed the `Provider`
type to a `Supplier` one: the lambda will be serialized, but the
only thing it will capture is the `outputDir` provider which is
fine, configuration-cache wise.

Now, let me take a pill.

Fixes #302

Lambdas must be serializable in order to be configuration cache
compatible.