Extended postprocessing for 2m temperature task

[mlangguth89]

In this issue, the postprocessing of models trained for the 2m temperature downscaling task will be extended. This includes:

• [x] Refactor code to make it more modular
• [x] Allow postprocessing from netCDF-file (instead of running inference on trained model)
• [x] Addition of the Mean Error of standard deviation-metric (see Eq. 6.4 in Zerenner et al., 2017)
• [x] Addition of conditional quantile plots
• [x] Block bootstrapping for uncertainty estimation
• [x] Plot examples

[mlangguth89]

The postprocessing can now be configured with a json-file. The content of this file should be as follows:

{
  "varname": "t2m",
  "unit": "K",
  "do_evaluation_time": true,
  "config_evaluation_time": {},
  "do_evaluation_spatial": true,
  "config_evaluation_spatial": {},
  "do_spectral_analysis": true,
  "config_spectral_analysis": {},
  "do_feature_importance": true,
  "config_feature_importance": {"score_name": "rmse"}
}