Thanks for the report.
The mri_with_eeg.py demo is known not to work (I just haven't got around to updating it yet).
As for speed, your results are consistent with mine. There is probably still some low hanging fruit to be optimized on the branch, but I expect that the new design will ultimately be slower than the current trunk, due to moving a number of things into Python/Numpy from pure C and the increased generalization of the code.
Cheers,
Mike