Fix vLLM wrapper option forwarding

[patrick-andstar]

Summary

• Forward the caller's quantization option to vLLM.
• Include caller-provided stop token IDs in vLLM sampling parameters.
• Store (query, response) pairs in returned history.

Motivation

The vLLM wrapper silently ignored supported call options and persisted the generated prompt instead of the user query, producing incorrect follow-up context. The same implementation exists in the example and recipe copies.

Testing

• Ran a stubbed vLLM regression harness against both wrapper copies.
• Ran AST syntax checks for both files.
• Ran git diff --check.
• Did not run real vLLM/model integration because it requires the vLLM runtime and a checkpoint.

Duplicate Work Check

Searched open Issues and PRs for vllm_wrapper, quantization, and stop_words_ids; no matching open change was found. Closed issue #975 reported related vLLM context behavior but did not change this implementation.