[Design] Setting Rework

[lebrice]

Leaving this as a note for now, will update with more info shortly.

Todo:

• Have settings/assumptions as "wrappers" around other settings: transition from a "is-a" (inheritance) to a "has-a" (composition) approach.

For example:

my_setting = TaskIncremental(
    SemiSupervised(
        Cifar10(data_dir="/data"),
        labeled_fraction=0.5,
    ),
    num_tasks=5,
)
# Ideally, have a custom __instancecheck__ for these wrappers, so that we can have this:
isinstance(my_setting, TaskIncremental) # True
isinstance(my_setting, SemiSupervised) # True
isinstance(my_setting, Cifar10) # True

• The ordering of these "wrapper functions" shouldn't have any effect.
• They would also be singledispatch callables, so that we can register handlers for different environments or for SL / RL.
• Ideally directly accept:
  • a Scenario from Continuum
  • a DataModule of some sort from pytorch-lightning-bolts
    and adapt them into the required kind of setting.