gha: test multiple versions

[dtrifiro]

• gha: test multiple vLLM versions
• tests: use real vLLM version (stop using mocks)
• grpc_server: only print memory summary for CUDA devices
• tests: add real tests (for both grpc and http server)
• enable code coverage

The reason for this:

1. We can test multiple vLLM version and catch any issues before they hit PyPi
2. We can test a real vLLM engine by using the CPU executor

[prashantgupta24]

https://github.com/opendatahub-io/vllm-tgis-adapter/blob/main/src/vllm_tgis_adapter/__main__.py#L251

image_input_type was removed here - vllm-project/vllm#5852

[dtrifiro]

Handling image input shape issue in #31