feat: add LLMTR passthrough model catalog

[knowhycodata]

Follow-up to #2062 (which added LLMTR + its first-party models).

LLMTR is an OpenAI-compatible gateway that also proxies models from other providers. This enumerates the passthrough models it serves so they show up in the catalog without per-user config — the same approach used for other gateways here.

How: adds packages/core/script/generate-llmtr.ts (bun llmtr:generate). It reads LLMTR's public /api/models endpoint, keeps CHAT_COMPLETIONS models, matches each to an existing models.dev base model, and writes base_model + LLMTR's price + served context. Reasoning models get reasoning_options = [{ type = "toggle" }] (LLMTR's chat gateway exposes a reasoning on/off toggle, not effort levels or token budgets — see the notes in provider.toml).

Result: 81 passthrough models across anthropic, openai, google, qwen (alibaba), zai (zhipuai), mistral, minimax, deepseek, perplexity, stepfun, moonshot.

Scope notes:

• Only CHAT_COMPLETIONS models (RESPONSES-only ones like gpt-5.5 / codex / grok don't work over the OpenAI-compatible chat path).
• Models with tiered pricing or no models.dev base are skipped by the generator (logged).
• Everything uses base_model, so capabilities/limits/modalities/description are inherited from the authoritative base entries.

bun validate passes (87 LLMTR models total). Verified with opencode: opencode run --model "llmtr/anthropic/claude-haiku-4.5" completes a chat.