Instead of producing a SimInit object, a bench is now expected to return
a fully constructed simulation with its scheduler.
This means that the client does not necessarily need to provide the
starting time for the simulation. This start time may be hardcoded in
the bench, or may be taken as a parameter for the bench configuration.
This change make it possible for benches to do more, for instance to
pre-schedule some events, or to do less, for instance by hardcoding the
simulation time rather than accept an arbitrary simulation time.
The API style is now more uniform: both are passed by mutable ref, and
only expose accessors. Additionally, the methods that were initially
accessed through the scheduler field are now directly implemented on
`Context`.
Previously, the scheduler key used the target model as subkey to order
messages that target the same model.
Now this subkey is the origin model rather than the target, or in the
case of the global scheduler, 0. This doesn't change anythin in practice
for the local scheduler since the origin and target models were the
same, but for the global scheduler this provides additional guarranties.
For instance, if the global scheduler is used to schedule an event
targetting model A and then an event targetting model B where the latter
triggers a message to A, it is now guarranteed that the first message
will get to A before the second.
Now that `step_by` returns an error anyway (it was unfaillible before),
there is no more incentive to keep it as a separate method.
The `step_until` method now accepts an `impl Deadline`, which covers
both cases (`Duration` and `MonotonicTime`).
The build context is now passed as a mutable reference due to the need
to mutate data when adding a model.
Contains small unrelated cleanups and documentation improvements too.
The external_input example has been as well adapted and (at least
temporarily) simplifiedi/modified to remove the dependencies on
`atomic_wait` and `mio`.
