Refactor tasks

[bengioe]

This PR refactors tasks into reusable components that will make it easier to define new tasks and methods.

The PR is pretty big, but a lot of it is moving code around (or deleting repeated code):

• Creates a notion of Conditional, implements TemperatureConditional, MultiObjectiveWeightedPreferences, FocusRegionConditional
• Separates the commonalities between seh_frag and seh_frag_moo into a more generic StandardOnlineTrainer that is meant to be easily subclassed for new tasks.
• Adds some implementation notes and comments
• Adds a validate_batch routine that's useful for debugging, e.g. new environments and datasets

Also fixes some bugs:

• Makes valid_offline_ratio a flag and sets it explicitly in tasks where it wasn't properly set
• first_graph_idx was incorrectly calculated in SubTB (affected logging of logZ values)
• QM9Dataset was returning the wrong shape for its rewards
• Adds a allow_5_valence_nitrogen flag to MolBuildingEnvContext, this is needed in some cases, see tasks/qm9.py.
• Adds an explicit stop_mask to MolBuildingEnvContext.graph_to_Data
• Fixes incorrect default objective name in seh_frag_moo
• Fixes the default configurations in the tasks' main that hadn't been updated
• Fixes a number of routines where focus_cond was assumed to exist (but we can now turn it off).

[hadim]

Not much to say!

Just a simple curiosity question below.

[maclandrol]

See comments.