Get automatically the index of the module redshifting for the parameter_list dictionary when getting the real redshift for computing the best model SED.

Modify ObservationsManagerPassbands to iterate over an instance of observationclass,  modify Workers accordingly and to include the properties and modify pdf_anaysis/utils to include properties in compute_chi2, _cimpute_scaling and dchi2_over_ds2

The standardisation of error bars must be done on the final estimates, not on the estimates for a single block.

Do not correct intensive physical properties for the difference in redshift between the observation and the grid.

Compute the best fit at the at the exact distance of the analysed object and not at the distance of the rounded redshift. This eliminates the need for a correction factor.

Give the redshift of the model rather than the list of redshifts to correct for the redshift difference between the model and the observation.

Rename ParametersHandler to ParametersManager to be consistent with new managers that will come in the near future.

Limit the number of threads to 1 per process in order to prevent MKL from creating numerous threads leading to a strong oversubscription of the CPU/RAM as pcigale already does its own parallelisation.

As we loop over the bands to compute the χ², it is smarter that the bands are the slower varying index and the models the faster varying index. That way it is more likely that models at the correct redshift are prefetched into the cache.

Better handle the case where one of the redshift is 0 to correct the physical properties for a possible offset between the observed and the model redshifts.

Correct the mass-dependent physical quantities for the difference between the redshift of the model and the redshift of the observation. Because the former is rounded to 2 decimal places, there can be a redshift difference of 0.005 (if CIGALE computes the list of redshifts itself). This can lead to differences of more than 0.35 dex at z=0.01 for instance.

When stripping the _log suffix of a variable names, let's just remove the last four characters instead of relying on strip(), which can have nasty side effects.

When estimating a parameter in log, these were not scaled appropriately and taken in log when saving the related PDF

When estimating a parameter in log, these were not scaled appropriately and taken in log when saving the related χ²

Get rid of the array containing the redshifts of all the models. Now that we infer the redshift of individual models from the list of unique redshifts it is not needed anymore. This change saves 8 bytes per model. The code should be marginally faster too.

Handle the case where we do not know the redshift to estimate the physical parameters by taking all models at all redshifts. This means that in effect pcigale can also act as a photo-z code now on.

Now that we leave the models with NaN fluxes, the scaling factors and χ² may be NaN too. Handle this case.

Using fancy indexing to select all the models as a given redshift is very suboptimal as it forces a copy for each of the index arrays. Now because the redshift is the last parameter, we can just use a proper slice. The first element is the index of the redshift in the list of redshift (paying attention that this list is in the same order as the computation order), and the step is the number of redshifts.

Because we are looping on the bands now and we can just test at each iteration whether it is a valid band, we do not need to create a new array without the invalid bands. This should save some memory bandwidth and time.

Now that we do not use magic numbers anymore but NaN values, we do not need to select the valid models. We can just compute on the entire grid. Invalid models will have a χ² of NaN and will therefore naturally and elegantly be eliminated from further analysis.

Magic values are a bit of a pain to handle and are not safe. This patch removes the use of magic values and rather replaces them with NaN whenever appropriate. The logic of the input file stays the same but the magic values are converted internally after reading it so that the rest of the code does not have to deal with that.

Computing the parameters and their uncertainties through the histogram of the PDF makes the code more complex and can introduce biases in some cases. This patch moves the computation of the histograms to the function that saves them onto the disk. In parallel the estimated value of the parameters and the corresponding uncertainties are simply computed from the weighted mean and standard deviation of the models that are at least 0.1% as likely as the best model to reproduce the observations.

To compute the reduced χ², we need to divide by the number of bands-1 (and not the number of bands). We do that because we consider that the models depend on one meta-parameter.