feat: add service_tier parameter to Responses API LLM

[piyush-gambhir]

Summary

The Chat Completions LLM (openai.LLM) already supports the service_tier parameter for configuring priority/flex/default processing per-request. The Responses API LLM (openai.responses.LLM) is missing this parameter despite the OpenAI Responses API supporting it.

This PR adds service_tier to the Responses LLM for parity.

Changes

livekit-plugins/livekit-plugins-openai/.../responses/llm.py (1 file, 6 lines):

• Add service_tier: NotGivenOr[str] to _LLMOptions
• Add service_tier parameter to LLM.__init__()
• Pass service_tier through in chat() via extra kwargs

Usage

from livekit.plugins.openai import responses

llm = responses.LLM(
    model="gpt-5.4",
    service_tier="priority",  # now supported
)

Backward Compatible

• Defaults to NOT_GIVEN — no impact on existing code
• Matches the existing pattern used by the Chat Completions LLM

[piyush-gambhir]

The CI failure is from test_blockguard.py::TestStress::test_many_short_blocks — a pre-existing flaky test in livekit-blockguard that's unrelated to this PR.

The test runs 20 × time.sleep(0.02) with a 500ms threshold and expects no blocking detection. On the CI runner (macOS), cumulative scheduling jitter causes the event loop heartbeat to exceed 500ms between watchdog polls, triggering a false positive. This happens because the test's total blocking time (20 × 20ms = 400ms) is close to the threshold, and any GC pause or scheduling delay pushes it over.

This PR only modifies livekit-plugins-openai/responses/llm.py — no relation to blockguard.

[piyush-gambhir]

Closing to re-trigger CI (flaky blockguard test failure unrelated to this change).