fix: normalise in place to reduce memory

[japols]

This PR avoids storing large intermediate results during the normalization by computing it in-place, e.g.

x[..., :] = x[..., :] * self._norm_mul + self._norm_add # old version x.mul_(self._norm_mul).add_(self._norm_add) # new version

Note that the in_place argument still works as expected, only the actual calculation is changed to be done in-place.

For 9km data the old version resulted in mini-peaks during the normalizer:

Version	Peak memory (GB, during normalization)
old	32.62
new	17.73

This lines up with storing two additional intermediate results (multiplication, addition) @~8GB batch size in the old version.