Introduce model input API

Model networks expect a set of parameters.

For example, DeepAR (mxnet) takes the following values:

https://github.com/awslabs/gluonts/blob/9260639dd28791e48db694d409fe5ef1eeb6806d/src/gluonts/mx/model/deepar/_network.py#L874-L886

It would be useful to make these parameters available on Estimator and Predictor to expose what input values (incl. type, etc) are expected.

We already do this to some extend in torch, where we expose input_shapes and input_types on some models.

Ideally we have single way of retrieving this information from all trainable models, both mxnet and torch.