Adding support for brainvision format

[hubertjb]

This PR adds support for dumping/loading mne.io.Raw objects using the BrainVision file format. As we still need to cache MEG data using .fif, the approach I took here is to identify the channel types before dumping, and to do some trial/error when loading.

TODO:

• Review file format selection logic at dumping and loading time @kingjr
• Organize the multiple files (.eeg, .vhdr, .vmrk) into a single directory per item
• Handle float64 warning when dumping to BrainVision format