Add balance_weights to weight balanced batches

[FrenchKrab]

The balance option of the segmentation tasks allows to pass a list of ProtocolFile fields, e.g. ['database', 'foo']. Then when batches are sampled, it looks at all existing combinations of values for these fields in the task protocol.

For example if they come from databases aishell and ami, and their foo field is either a or b, we compute the cartesian product [('aishell', 'a'), ('aishell', 'b'), ('ami', 'a'), ('ami', 'b')], batches are created by randomly selecting one of these tuples and picking a sample from a matching file.

The PR allows to weight the random choice from the cartesian product. For example with

balance_weights = {
  ('aishell'):2.0,
  ('ami', 'b'): 4.0,
}

we will sample from the cartesian product using random.choices with these weights:

selected = random.choices(
    population=[('aishell', 'a'), ('aishell', 'b'), ('ami', 'a'), ('ami', 'b')],
    weights=[2.0, 2.0, 1.0, 4.0],
    k=1,
)[0]

e.g. for each tuple of the cartesian product, we find the longest matching (tuple) prefix in balance_weights and use this weight.

I'm not sure this approach is flexible/clean enough to be PR-ready, and it's hard to make the docstring concise, but i think it could be really useful :)

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[FrenchKrab]

I wiped out the old branch since it was made too many commits ago. This is my updated proposal.
The task now have the balance parameter typed as :

balance: TaskBalancingSpecifications | Sequence[str] | dict | None = None

TaskBalancingSpecifications is the new class that handles all logic related to weighting/balancing rules. Sequence[str] is to keep compatibility with old code (uniform balancing).
dict is redundant, but it allows the user to easily specify balancing rules easily when using hydra (without having to instantiate the TaskBalancingSpecifications class)... maybe it is not necessary?

[FrenchKrab]

NOTE: it also implements #1787, which should be closed it this is merged

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.