refactor imputation transformation

[benidis]

Small refactoring of AddObservedValueIndicator transformation.

Removed unnecessary fields:

• dummy_value: This is imputation-method-dependent and should not be an argument of the transformation. Instead the transformation should take as input the imputation method with this field set already.
• convert_nans: This is unnecessary since we can just use None as imputation method.

With the above change we have control IF we want to do imputation or not in the transformation by including a hyperparameter in the estimator (setting it to the appropriate imputation method or to None). This gives the flexibility to select from a model-based imputation method (e.g. if set to None use a sample from the predictive distribution since the imputation from the transformation is disabled) or from a transformation based imputation (e.g. a dummy value).

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[codecov-commenter]

Codecov Report

Merging #907 into master will decrease coverage by 0.00%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #907      +/-   ##
==========================================
- Coverage   85.80%   85.80%   -0.01%     
==========================================
  Files         195      195              
  Lines       11853    11851       -2     
==========================================
- Hits        10171    10169       -2     
  Misses       1682     1682              

Impacted Files	Coverage Δ
src/gluonts/transform/feature.py	98.41% <100.00%> (-0.02%)	⬇️