cuda: float support for gpu transforms, contributed by Zekun Ni, see https://bitbucket.org/nschaeff/shtns/issues/44/float-single-precision-support

fix accuracy issue for lmax>1000 and grid including poles -- a full and clean fix, unlike commit 57666bf3

Fix shtns_rotation_set_angles_ZXZ() and shtns_rotation_set_angle_axis(). shtns_set_angles_ZYZ() was correct.

Merge bug fix for bitbucket issue #42 affecting complex-valued synthesis since commit d544d31b (v3.4)