Model blocks should be able to use the granite-io platform

[esnible]

Is your feature request related to a problem? Please describe.
We might be able to get performance and functionality improvements more quickly by implementing a back end that calls granite-io directly, rather than through LiteLLM.

Describe the solution you'd like
A back end that calls granite-io directly, rather than through LiteLLM.

Describe alternatives you've considered
Waiting for LiteLLM to incorporate all of the improvements offered by granite-io in a provider-agnostic way, and making use of those improvements.

[mandel]

This is how this PR is implemented.