Refactor EPP to decouple core logic from `InferenceModel` resource

[zetxqx]

As stated in the discussion on pull request #961, the current EPP implementation has a hard dependency on the InferenceModel resource. This coupling was identified during the implementation of the GatewayFollowingEPPRouting conformance test, where the testing EPP could not function without a concrete InferenceModel.

The current EPP logic requires an InferenceModel for two main reasons I spotted:

• To discover the universe of possible backend pods from the referenced InferencePool (code.
• To identify the modelName it is serving (code).