Round values from reprojected data

[remi-braun]

There is currently a blind spot with reflectance values coming from reprojected data.

For example, the loading workflow is usually:

Read (in float) > Reproject (bilinear) > Scale > Convert to float32

The issue with this workflow is that the bilinear reprojection (with float data) creates values with more precision than intended (i.e. for Sentinel-2, the precision is 10^-5).
This could introduce biases, especially when computing spectral indices.

It may exist a need of the community to round those values (see #103). If it does, this modification shouldn't be too complicated.