fix merge_asof usage in read_raw_eyelink()

[dominikwelke]

closes #12480

the new parameter set is also not perfect:
direction=nearest can lead to duplicating samples (as a sample on an uneven ms is equally distanced to the previous and subsequent millisecond). you can see this behavior running the simplified example code in the issue.

but i think this is better than forward/backward which does not duplicate samples, but works worse with sfreqs < 500 Hz, as it can shift samples in the wrong direction

[scott-huberty]

Thanks again! - I haven't looked at the code yet, but if merge_asof is still giving us some trouble I'm just going to crossref to my comment in the associated issue about refactoring the code to more closely the pattern in read_raw_neuralynx:

#12480 (comment)

[larsoner]

Thanks @dominikwelke