test: add inference replay harness MVP

[bk-nvidia]

Summary

• add a strict v1 synthetic replay fixture for the streaming 429 follow-up regression
• launch the real nemo-relay process against a raw-TCP local provider
• verify the 429 response and a successful follow-up stream on one reused upstream connection
• add deterministic deadlines, sanitized diagnostics, fixture validation, and a focused just command

This is the smallest Phase 1 slice. It replays synthetic workload structure; ATOF/ATIF compilation is intentionally left for a follow-up.

Validation

• cargo test -p nemo-relay-cli --test inference_replay_tests -- --nocapture
• cargo test -p nemo-relay-cli --test inference_replay_tests streaming_429_followup
• cargo clippy -p nemo-relay-cli --test inference_replay_tests -- -D warnings
• cargo fmt --all -- --check