Adding CrossTalk, making Choose better.

[pseeth]

This adds two things:

• CrossTalk transform: mixes items in the same batch with other items from that batch
• Changes how the Choose transform works. Choose now manipulates masks for the transforms so that only one is applied. Before the one transform that was applied was applied to the entire batch, whereas now, the one transform that is applied is chosen per item in the batch. Thanks @oreillyp for this change!

And then some fixes:

• hanning got removed from scipy 1.9.0, as it was just an alias for hann. So that broke some code, I took out all references to hanning and replaced them with hann in the codebase.
• The profile_transforms script needed some tweaks to accommodate GlobalVolumeNorm, as well as excluding Repeat and RepeatUpTo.
• Because of the changes to Choose, the regression data for RepeatUpTo had to be fixed.

[oreillyp]

Looks good to me. This may not be within the scope of this PR, but there are cases in which it may be beneficial to fix a transform across all batch indices (e.g. have Choose pick the same effect for each batch item, or have RoomImpulseResponse apply the same impulse response to all batch items). This could be accomplished by a constructor argument in BaseTransform that, if True, would result in all states given to .batch_instantiate() being deterministically replaced with the first state in the batch. This would allow toggling batch-randomness on/off for individual transforms, including Choose.

class BaseTransform:
    def __init__(self, keys: list = [], name: str = None, prob: float = 1.0, batch_randomize=True):

    ...

    def batch_instantiate(
        self,
        states: list,
        signal: AudioSignal = None,
    ):
        if not self.batch_randomize:
            states = states[0] * len(states)

        kwargs = []
        for state in states:
            kwargs.append(self.instantiate(state, signal))
        kwargs = util.collate(kwargs)
        return kwargs

I believe this means the modified states would then propagate down the hierarchy of any composed transforms within the transform in question, which seems like reasonable behavior. This way, all future transforms could be written with the new Choose approach of randomizing across batch indices, with the constructor argument giving the option to fix transforms across batch indices.

[pseeth]

Hmm, I think it's a good idea but should be a separate PR. batch_instantiate is a helper function, but if you look in CSVDataset, the transforms are actually instantiated on separate threads, and then collated together afterwards.