(feat) Introduce Router (litellm) into LLM and LLMConfig classes

[tobitege]

Short description of the problem this fixes or functionality that this introduces. This may be used for the CHANGELOG

Introduce Router and configuration (litellm) into LLM and LLMConfig classes. Resolves #4056

---

Give a summary of what the PR does, explaining any non-trivial design decisions

• Initial PR to add configuration and use of litellm's Router class
• In LLM class regular completion (as before), only use router if router config incl. models exist
• Added extensive example router configuration to config.template.toml with multiple models
• The load_from_toml method loads section router_config as a top-tier section into apps' llms["llm"] The router's configuration is complex as it is, hanging it in toml syntax under main llm in the toml file would make this even more complex and potentially unreadable, due to nesting, imho
• Added new test_llm_router.py unit test file with router-related tests

Other thoughts

• I see this as an add-on, not as a replacement for the existing completion logic, as it requires a config.toml file and a rather complex configuration.
• Its only used if such configuration exists in the config.toml file. Otherwise the regular completion mechanic is used as before.
• The router does retries internally, the specifics need to be looked up in the docs.
• From their docs: For RateLimitError we implement exponential backoffs

---

Link of any specific issues this addresses

litellm Routing documentation