[ENH] refactoring varying integration test parameters to v1 test framework

[fkiraly]

For the purpose of the v1 testing framework, the varying parameters in estimator specific tests should be moved to the get_test_train_params method of the metadata classes.

For this, we should check:

• in the _integration fixtures, which variables vary
• ensure these are transferred to get_test_train_params as explicit dict keys - for all algorithms - so the TestAllEstimators run agrees 1:1 with the individual _integration runs

Parameters identified so far that are not explicitly in kwargs of all fixtures:

• learning_rate (in from_dataset call), this varies across networks
• cell_type, n_plotting_samples (in from_dataset call) from DeepAR
• backcast_loss_ratio (in from_dataset call) from NBeats

[fkiraly]

FYI @phoeenniixx, @PranavBhatP, @fnhirwa