Feature/preparation for 0.13

- add, prune and exclude docs where appropriate
- .rst fixes

update the documentation and include it in the README

this is relevent for constants, which we usually do not want to be documented
separately since they are displayed as default values for __init__() anyways

refactor filterbank arguments to FilterbankProcessor; closes #92 

Remove the option to normalize the observations of the ObservationModels.

Made some properties of MIDIFile methods; fixes #85 

data stream

loading

Make audio.signal package more consistent

frames to left of the first frame can still be accessed with frame_signal()

normalize() rescales the signal to have maximum range for the dtype
remix() converts to float and back if needed to prevent clipping when down-mixing

numerous fixes for v0.13

Use uint32_t for dtype=np.uint32 array.

On some systems unsigned int are interpreted as a 64-bit unsigned int instead of a 32-bit.

Refactor the way the spectrogram differences are stacked. 

SpectrogramDifferenceProcessor accepts a numpy stacking function as its
`stack_diffs` argument
SpectrogramDifference keeps a reference to the underlying spectrogram instead
of the STFT
deprecated StackedSpectrogramProcessor

remove pickling of data classes; fixes #79