Introduce a non-global Vertex AI client

Currently, vertexai.init() sets global variables which is a limit in a multithreaded environment when you want to serve different models to the corresponding available regions. Having an instantiable client (as in your nodejs [client](https://cloud.google.com/vertex-ai/generative-ai/docs/reference/nodejs/latest/vertexai/vertexai) would solve the problem)

Start with the vertexai.init() entry point and compare its global-state behavior with the linked Node.js Vertex AI client. Define the instantiable Python client’s scope and verify that separate clients can serve different models in different regions without shared configuration.

Written by the indexing model from the issue text.