Add `DurableCacheEvaluator`

Extend the existing in-memory caching prototype with a durable backend so that expensive CallableModel results survive process restarts. Long-running ETL, training, and reporting workflows can then resume without redoing completed work, and repeated runs across sessions reuse prior outputs.

Semantically this is the same as the in-memory cache: a write-through, invalidatable optimization. A miss simply triggers recomputation. It composes naturally with the in-memory cache as an L1/L2 layer.

Behavior

• Pluggable storage backend (diskcache is a reasonable default; object storage is a useful follow-on).
• Same identity / cache-key semantics as the in-memory evaluator.
• Standard cache controls: TTL, size limits, manual invalidation.

Open questions

• Identity and staleness. What goes into the cache key beyond model identity and context? Source hashing, explicit user-bumped versions, or a combination?
• Large results. Inline storage of multi-GB results is impractical. Should result types be able to spill to external storage and store only a pointer in the cache record, and how does that interact with the arrow/narwhals/pandas result types?

Start by reading the existing in-memory caching prototype and its identity and cache-key behavior. Resolve the open questions around staleness, durable storage, large results, TTL, size limits, and invalidation before implementing the evaluator. Done means results survive process restarts while preserving the in-memory cache semantics and supporting an L1/L2 composition.