Fewer fast floor impls for image interpolators

There were several checks for specific CPU architectures, many
of which are of only historical interest.  The new code has only
three implementations: one for 64-bit CPUs, one for little-endian
32-bit CPUs, and one for big-endian 32-bit CPUs (the latter two
implementations are identical apart from the endianness).

The previous fallaback that used vtkMath::Floor() would give the
same integer from floor, but a slightly different fractional
remainder.  The difference was at most 2^(-16), and therefore not
generally significant, but the this fallback was superfluous
since the three implementations described above are sufficient.