Multiplicative noise: add possibility to use maps of standard deviation instead of only one general value

[pierrePalud]

As of today, the standard deviation (std) of the multiplicative lognormal noise is set in the input_params.yaml, in the parameter

...
sigma_m_float_linscale: 1.3
...

So that \sigma_m = \ln ( sigma_m_float_linscale ). In the example, 1\sigma error corresponds to an error by a factor of 1.3 (that is, about 30%). The multiplicative noise gathers both error calibration and model mis-specification. In the current implementation, the sigma_m_float_linscale value is set for all observables (i.e., lines) and all pixels. This is problematic for both calibration error and model mis-specification.

Difference between observables:

• For calibration error: the expected calibration error might change when the observables come from different instruments. Similarly, for one telescope, the expected calibration error might change with the frequency band (see e.g., https://www.aanda.org/articles/aa/full_html/2024/11/aa51588-24/aa51588-24.html ).
• For model mis-specification: in numerical models, some observables are considered better modelled than others. The inversion could exploit a lower multiplicative std for these observables.

Besides, for a given observable:

• a region where the considered numerical model is less trusted should have a higher std.
• a region that has been observed more times than others should have a lower calibration error and thus multiplicative noise std.

Therefore, for generality, it is important to give to the user the possibility to set a map of std values of the multiplicative noise.

• give to the user the possibility to provide a map of std for the multiplicative lognormal noise, in the same way as the std of the additive Gaussian noise.
• adapt documentation.