Benchmark UDS vs embedded tokenizers

[sagearc]

As discussed in the comment here: #295 (comment), we should run a benchmark comparing the embedded tokenizer and the UDS tokenizer. Ideally, this should be done without the indexing logic involved.

[hyeongyun0916]

I think we'll get more accurate results if we:

1. Run the benchmark after merging uds tokenizer using vllm wrapper #280.
2. Compare only render_chat / render performance, excluding the get_or_create_tokenizer_key(init) overhead.

This should help us isolate the actual tokenization/rendering latency.

[sagearc]

ref: llm-d/llm-d#722