Enable training with streaming harnesses

[hallerite]

Summary

• bump the verifiers submodule to the focused streaming-training bridge in verifiers #2369 at signed commit 80250d62b
• allow renderer-backed training clients to satisfy streaming-only Chat Completions harnesses without requiring renderer token streaming
• start generation asynchronously so the verifiers interception server can send SSE headers and keepalive comments while the buffered generation is still running
• emit the completed response as synthesized SSE, while committing the original typed response with exact token IDs and logprobs
• admit verifiers' pinned Harbor 0.21.0 through prime-rl's rolling dependency cutoff and refresh the lockfile (including the resulting LiteLLM update)

Validation

• uv lock --check
• verifiers: uv run ruff check --fix .
• verifiers: uv run ruff format --check .
• verifiers: uv run pytest tests/v1 -m 'not e2e' — 70 passed
• verifiers: uv run pytest tests/ — 914 passed, 75 API-key e2e tests skipped
• delayed-generation smoke: relay returned before generation completed, leaving the server able to emit keepalives; synthesized SSE then parsed successfully and retained the generated TurnTokens
• prime-rl renderer-client configuration smoke against the bumped verifiers source

This deliberately does not add token-by-token renderer streaming. The harness sees a valid streaming response and connection liveness, while training still consumes the exact buffered renderer result.