Add Hopper LLM provider

[pavanyellow]

Summary

Adds a LLM.with_hopper() factory to the OpenAI plugin, alongside the other
OpenAI-compatible providers (Cerebras, Together, Nebius, Telnyx, …).

Hopper serves open-source models optimized for low
time-to-first-token, aimed at voice agents. The API is OpenAI-compatible, so
this follows the existing with_* provider pattern — no new plugin, just a
factory method plus a HopperChatModels type.

Usage

from livekit.plugins import openai

llm = openai.LLM.with_hopper(
    # api_key defaults to HOPPER_API_KEY env var
    model="Qwen/Qwen3.6-35B-A3B",
)

Get an API key at https://withhopper.com.

Changes

• with_hopper() classmethod on openai.LLM (defaults: base_url=https://api.withhopper.com/v1, HOPPER_API_KEY).
• HopperChatModels literal in models.py.
• _strict_tool_schema=False, matching the other open-model providers served on vLLM-style backends (same as Cerebras, use non-strict tool schema for cerebras llm #3134).

Latency

TTFT (time to first token) over a warm connection, voice-agent-shaped context
(~2k-token system prompt + short user turn), 10-run p50:

• From us-west-2 (same region as the model server): p50 62ms (min 56, max 77)
• From a residential laptop in SF: ~170ms — the difference is network round-trip

First-token latency is dominated by where your agent runs relative to the model,
not the serving itself; colocated, it's ~60ms.

Testing

Verified against the live endpoint through the plugin:

• factory builds with the correct base URL and model
• streaming chat() returns a valid completion
• function calling emits a correct tool call (validates _strict_tool_schema=False)
• ruff check passes on the changed files

[CLAassistant]

All committers have signed the CLA.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.