MultiTypeBlockVector: Fix NaN behaviour of infinity_norm()

As a follow-up to flyspray/FS#1147 (closed), in response to dd7e8ed3, this makes MultiTypeBlockVector::infinity_norm() NaN-aware: In analogy to previous patches for other matrix/vector classes, this patch turns infinity_norm into two functions; a straight-forward version for field_types that do not have NaN values and a less efficient version for ones that do.

It should be noted, however, that because field_type is currently hard-coded to be double in MultiTypeBlockVector, the NaN-aware version will, as long as this is so, always be used.