Support BDF files

[ararslan]

BDF is BioSemi's variant of EDF that encodes samples as 24 bits each rather than 16. Here we're using BitIntegers.jl to get an Int24 type that Should™ allow us to support BDF with minimal internal changes.

Closes #33.

Current state of this PR: Works! Doesn't fully work, haven't looked too closely yet, was mostly just trying this out for fun.

[klaff]

I might be missing something here. Why use Int24 as opposed to just using Int32? I don't think you save on memory due to alignment considerations (i.e. sizeof(zeros(Int24, 100)) and sizeof(zeros(Int32, 100)) both give 400) and I don't think we would be relying on things like 24-bit specific rollover characteristics.

[palday]

@klaff I think the problem comes when moving it to/from disk: you need to interpret the blitted data as Int24 and also blit it back out as Int24. Once it's in memory though .....

[klaff]

Ah, so some bit-twiddling like:

function convert_24b_bytes_to_Int32(MSB, midB, LSB)
    return reinterpret(Int32, [0x00, LSB, midB, MSB])[1] >> 8
end

[ararslan]

I think we've generally tried to keep the internal representations of things as close as possible to how the format stores them, for better or for worse. (One notable deviation is that the array of samples for each signal is one big vector and is not divided into data records. Annotations, however, are divided into data records.) Keeping the encoded representation true to the format also simplifies both the reading and writing; no extra conversions or checks are required. For actually using the data, you'd want to call EDF.decode anyway, which will give you a Float32 array for both EDF and BDF files.

[ararslan]

Note that this requirement had to be removed because the version field of a BDF header begins with a non-ASCII byte.