[ENH] Speed up read_raw_neuralynx() on large datasets with many gaps

[KristijanArmeni]

Reference issue

Fixes #12370.

What does this implement/fix?

Drop np.where and compute onsets via np.cumsum Instead of using np.where to iterative over gap indices and search over the recording samples, in this PR I compute all segment start samples by cumulatively summing segment sample sizes (see this change: 5b40d2).

This gives starting indices for all segments, then I select only those corresponding to gaps. Doing it this way, yields substantial improvements over using np.where as confirmed on a a few local datasets with increasing numbers of gaps (see Figure in #12370)

Additional information

Any additional information you think is important.

[larsoner]

I don't think a changelog entry is needed for this so I'll label it as such, but feel free to add one if you want @KristijanArmeni (probably as 12371.newfeature.rst since this is an enhancement) . I'll mark for merge-when-green, thanks in advance!

[KristijanArmeni]

I don't think a changelog entry is needed for this so I'll label it as such, but feel free to add one if you want @KristijanArmeni (probably as 12371.newfeature.rst since this is an enhancement)

Thanks @larsoner! Added a changelog entry as suggested.

[larsoner]

pip-pre failures are unrelated, I opened #12372 to track it

[KristijanArmeni]

@larsoner just pinging here that it looks like this was not auto-merged (it was automatically disabled after my changelog commit). But I see pip-pre fails are now sorted and main is passing so I think just this just needs an update from main and should be all green.

[larsoner]

Thanks for the ping @KristijanArmeni, merged main into this branch reenabled auto-merge, and will keep an eye on it to make sure it lands today!