consider replace double by float to speed up

[djeanner]

We discussed whether to use float instead of double.
For fft, it may not be a good idea (because of variable multiplied by values large numbers of times). For the all other operation (symmetrization, scalar product, deconvolution per se) both are fine.