two bug fixes (should not have caused any problems yet)

This fixes two problems:

• the requested size in aligned_alloc must be a multiple of the alignment, which we didn't enforce. This apparently hasn't caused any problems so far, but better safe than sorry.
• r2r_fftpack used the wrong flag for the direction of the underlying halfcomplex transform. This is only a concern in cases where real2hermitian!=forward, and this scenario was not yet used, at least not from scipy.

@g-peterbell: I guess this should go into scipy at some point, but it's probably not urgent. What do you think?