ds.start() unexpectedly loses input files

[lukemarsden]

a very common use case is that you do ...

ds.input(read_in_some_data())

and later...

ds.start()
model.fit()
ds.summary("x_accuracy", model.accuracy)
ds.publish("trained a model")

ds.start()
model2.fit()
ds.summary("y_accuracy", model.accuracy)
ds.publish("trained some other model")

i think we just shouldn't reset input files on ds.start() and instead have an explicit ds.clear()

we should clear params & stats on ds.start() though

[lukemarsden]

We could also make ds.start() only clear things if a previous run has happened. Might this be an improvement, or does it make the behaviour confusing?