The goal is to rewrite the KServe v2 protobuf parser (among other things, needed for Phase 1 and "real" feature names).
With the new parser, the following KServe payloads are supported:

• PD, single/multi-input and single/multi-output, single request and batch request
• NP, single/multi-input and single/multi-output, single request and batch request

A new datastructure (TensorDataframe) will also be added, which extends Dataframes and allows to convert existing prediction inputs and outputs to the correct PD/NP format, either row-wise and as a batch for the whole dataframe.

Acceptance criteria:

• Both single and batch requests payloads should be saved in storage as expected.
• Dataframe-like payloads will have the correct input and output names, and for array-like, feature names can be provided externally (as they do not exist in the model’s metadata).
• Correct creation of infer requests and responses from a TensorDataframe

The unit tests cover the direct conversions themselves, the tensor/dataframe conversion and for the TrustyAI service consumer endpoints.

Upstream issue: https://github.com/trustyai-explainability/trustyai-explainability/issues/153