Warn when reading discontinuous EDF+ files

[palday]

Due to the way EDF+ implements the discontinuous extension, a naive EDF reader can still open the file but treats the signals as continuous. This is a potential footgun -- I only noticed it because the annotations referred to times that were much later than the concatenated signals referred to because the annotations assumed that the discontinuity was correctly handled.

[ericphanson]

we could use this check:

function is_functionally_contiguous_after_start(file::EDF.File)
    # EDF+ says: first annotations signal is such that the first
    # annotation in each list is the start time of the record
    time_idx = findfirst(x -> isa(x, EDF.AnnotationsSignal), file.signals)
    time_anns = file.signals[time_idx]
    starts = [anns[1].onset_in_seconds for anns in time_anns.records]
    # if the start of the next record is exactly seconds_per_record + the start of the previous record
    # then the file is "functionally" contiguous
    return all(==(file.header.seconds_per_record), diff(starts))
end

When this holds, the start may not be at the header's start, but after that all the records are contiguous and reading works. To get the signal start, you need

function start_offset(edf::EDF.File)
    i = findfirst(s -> s isa EDF.AnnotationsSignal, edf.signals)
    i === nothing && return Nanosecond(0)
    annos = edf.signals[i]
    # Shouldn't happen per the spec but who knows, nobody seems to care about the spec
    (isempty(annos.records) || isempty(first(annos.records))) && return Nanosecond(0)
    tal = first(first(annos.records))
    return Nanosecond(tal.onset_in_seconds * 1e9)  # convert to ns since may be non-integer
end

Then when mapping to Onda format, the header's start corresponds to the recording start, and the start_offset corresponds to the signal span's start.