Add support for stop_words in Ray MBridge deployment

[athitten]

Extracts stop_words from incoming request and exposes them in nemo_deploy/llm/megatronllm_deployable_ray.py and nemo_deploy/llm/megatronllm_deployable.py to be passed along to the mcore inference engine. Helps reduce unnecessary token generation(which was the case before where a lot of unnecessary tokens were generated) beyond the stop_words passed in the incoming eval requests hence improving the speed.

Speed improvement with 10% gsm8k eval on llama 3.2 1B:

Before stop_words support in deployment: 10 mins
After stop_words support in deployment: 4 min 37s

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[athitten]

/ok to test 7937894

[athitten]

/ok to test 1fedc31