Eliminate AbstractTaskTrialSpec

[mlprt]

Currently, typing in feedbax.task is a mess. A source of this mess is AbstractTaskTrialSpec.

feedbax/feedbax/task.py

Lines 150 to 167 in 1c239e6

	class AbstractTaskTrialSpec(Module):
	"""Abstract base class for trial specifications provided by a task.
	Attributes:
	inits: A mapping from `lambdas` that select model substates to be
	initialized, to substates to initialize them with.
	inputs: A PyTree of inputs to the model.
	target: A PyTree of target states.
	intervene: A mapping from unique intervenor names, to per-trial
	intervention parameters.
	"""
	inits: AbstractVar[WhereDict]
	# inits: OrderedDict[Callable[[AbstractState], PyTree[Array]],
	# PyTree[Array]]
	inputs: AbstractVar[AbstractTaskInputs]
	target: AbstractVar[PyTree[Array]]
	intervene: AbstractVar[Mapping[str, Array]]

Once (#10) the target field is specified in a generalized way similarly to inits, the only field that will vary systematically between tasks will be inputs. In that case I suspect it will make more sense to eliminate subclassing of AbstractTaskTrialSpec, and instead define it as a generic final class, something like:

InputT = TypeVar("InputT", Module, Array)

class TaskTrialSpec(Module, Generic[InputsT]):
    inits: WhereDict
    inputs: InputsT
    targets: WhereDict
    intervene: Mapping[str, Array]

This would also save developers from explicitly including the intervene field in subclasses of AbstractTaskTrialSpec, because subclassing would be unnecessary; see #20.

I am still not sure how (if at all) to explicitly associate the structure of a given type of inputs, and the type of model that is compatible with a task. In principle, it could involve multiple fields of TaskTrialSpec; see #14.

[mlprt]

AbstractTaskTrialSpec has been replaced by TaskTrialSpec as of 4f735f4.