Use results to create th uncertainties with pdf errors.

[scarlehoff]

This is necessary to allow for things like #2049 but in general it also cleans up the code as more things are concentrated under results.

@achiefa you will need to rebase on top of this branch if you want to add pdf errors but I think for the time being you don't need them so you can wait until it is in master.

[RoyStegeman]

Could you be more explicit in your explanation? What is the example nb trying to do that needs this? It seems that all this does is replace the exp covmat with an identity matrix, but I don't believe it is ever read anyway. Is it to save the computational resources?

[scarlehoff]

pdferr_plus_dataset_inputs_covmat needs to compute the predictions for a dataset in order to generate the covariance matrix with exp + pdf error.

Ideally we want to get everything in validphys through a result object* but this creates a chicken-and-egg problem because the results object includes the covmat, so this adds as well a result obect without a covmat.

*and the notebook is based on monkeypatching the results (and results_with_thcovmat) objects