fix(openrouter): use model context length

[snvtac]

Summary

• use OpenRouter model-level context_length for synced limit.context
• keep top_provider.max_completion_tokens as the output limit source
• update the current xiaomi/mimo-v2.5 OpenRouter TOML by removing the stale 32K context override
• add regression coverage for shorter top_provider.context_length values

Fixes #2796.

Validation

• bun test packages/core/test/sync.test.ts
• bun run test from packages/sdk
• bun validate
• git diff --check
• bun models:sync openrouter --dry-run / bun models:sync openrouter showed the target xiaomi/mimo-v2.5.toml data fix; unrelated live OpenRouter catalog drift was not included

Known existing failure

• bun test still fails catalog generation > repository open-weight model metadata includes weights links because existing model metadata entries are missing weights links (for example nvidia/nemotron-3-super-120b-a12b and related entries).