Fix EpochsTFRArray default drop log initialisation

[tsbinns]

Reference issue (if any)

Fixes #13027

What does this implement/fix?

Expands the logic for EpochsTFRArray default drop_log initialisation so __getitem__() works with default params.

[drammock]

scratching my head about this line. When would we not want just range(len(self.events))?

[larsoner]

We always have len(self.events) == len(epochs). But let's say you originally had 100 events and just selected the last 10 on way or another, that resulting Epochs object shouldhave for example

len(epochs) == 10
len(epochs.selection) == 10
len(epochs.events) == 10
list(epochs.selection) == list(range(90, 100))

and in general for anything not in epochs.selection you should be able to query that index in epochs.drop_log to see why it's not part of epochs. So epochs.drop_log[30] for example should exist and be non-empty meaning it was dropped for a reason, whereas epochs.drop_log[90:] should exist and all be empty (i.e., those are the epochs that have been kept).

[drammock]

right, I forgot about events getting truncated/sliced/subselected.

Am I right then that we don't need to care about the case of having 30 events, and selecting only the middle 10? IIUC, if that happened via normal means, the drop_log would be present and the code in this diff won't be reached. It's only for EpochsTFRArray (where we don't know the dropping history) where we have to spawn a "fake" drop log, and in that case we can't know if (say) the last 10 epochs were dropped, but we also don't need to know that for things to work correctly.

[larsoner]

Yeah I think we just need some drop log that satisfies the necessary expectations about length of the drop log, selection values, and number of events.

[tsbinns]

So something like

tuple(() if k in self.selection else ("IGNORED",) for k in len(self.events))

to account for the fact that a non-default selection param could be passed?

Or just the super simple

tuple(() for k in self.events)

[larsoner]

No the len(self.events) won't be enough, I think what you have here with the max(...) is more correct

[larsoner]

... or you have to make sure self.selection = np.arange(len(self.events))

[tsbinns]

Ah sorry. Yeah at the moment whatever selection can be passed so self.selection = np.arange(len(self.events)) can't be assumed.

[tsbinns]

Have now switched from as_array to as_tfr_array in the test, and have included a changelog entry.