[AutoAgent] Add Validation to AutoKernelAgent routed Models

[Jack-Khuu]

AutoKernelAgent does not validate whether the generated or cached router_cache.json entry is requesting models that are unavailable to the caller.

This PR introduces _validate_cfg_models which checks model availability against the provider when writing or loading from router_cache.json

---

Along the way, we also fix a routing bug where route_cfg is not properly set when there is a cache miss (i.e when router_model generates a new entry).

• This presents as a subtle bug where the first time a problem is used in Fuser.auto_agent the routing config is generated, but not utilized (instead CLI args are used). When the model is called a second time (with a cache hit), the config is then properly utilized resulting in different outputs (first one is wrong, second is right)

route_cfg = cache[code_hash].get("config") or {}

---

Test

• Added pytest tests/fuser/test_auto_agent_cache.py

• Checking with empty cache python -m Fuser.auto_agent --problem /home/jackkhuu/KernelBench/KernelBench/level3/33_VanillaRNN.py

• Manual testing with router_cache.json entry and observing that the cli model is used instead

  {'fuser_dispatch_jobs': 'auto', 'compose_max_iters': 3, 'fuser_verify': True, 'llm_models': {'extract': 'gpt-4o', 'dispatch': 'gpt-4o', 'compose': 'gpt-4o'}