FeatureRequest: Support more flexible input data

[mhuen]

The dnn_reco framework currently only has convolutional models implemented by default. The modular setup of the framework is however designed to allow for arbitrary architectures. The current limitation is the expected format of the input data, which is expected to consist of constant shape tensors. This needs to be made more flexible, for instance to also allow dynamic and ragged tensors for a pulse series of arbitrary length.