Add SGLang backend adapter

[algebra-MCX]

Summary

• Split the session runner backend layer into reusable backend adapters.
• Add native SGLang /generate support with token-id input, streaming response parsing, generated token-id recovery, and cache metric normalization.
• Keep the existing vLLM backend behavior through the new shared backend client.
• Add a llama.cpp placeholder so the next backend can be implemented without changing the runner architecture.

Validation

• cargo test --manifest-path replay/Cargo.toml
• Real GPU SGLang replay test: 10 sessions, 80 rounds, 80/80 successful, 18,684 generated tokens.
• Real GPU vLLM replay test: 10 sessions, 80 rounds, 80/80 successful, 18,684 generated tokens.