Calculate Accurate Prompt Tokens for Chat Completions in vLLM Client

[vivekk16]

Identify and resolve discrepancies in prompt token calculation for chat completion requests in the vLLM client. Currently, there's a consistent mismatch between the actual usage reported by vLLM server and estimated token count (~100–150 additional tokens) through custom tokenizer.

This issue will track improvements separately from the PR #43 to avoid scope creep.

[vivekk16]

Kindly assign this to me.
@SachinVarghese I'd appreciate your inputs on this implementation.

[SachinVarghese]

Sounds good @vivekk16

One of the other ideas here would be to convert the current tokenizer component to a more specific token counter component. This way we can have different kinds of token counters - based on API response and configuration - and simplify the vllm client code that can be extended to other model clients. As an example,

class CustomTokenCounter:
    def __init__(self):
        pass

    def count_tokens(self, api_type: APIType, input:str, output:str) -> (input_token_count int, output_token_count int):
        pass