BUG: Raise TypeError when converting mixed-type DataFrames, use proper NA types, close #1282

[lbeltrame]

This pull request attempts to fix two problems:

1. The use of improper NA types in R, which can cause issues when converting to matrices (NA_Logical gets converted to 0)
2. Improper casting when converting a mixed-type DataFrame to matrix, which will cause all sorts of problems: thus mixed types DataFrame raise a TypeError exception now.

Added a bit to the unit tests so that I know the thing works, but these (for pandas 0.9) needs to be ported to unittest/nose so that it can be run with the other tests.
#1282