Cache compiled artifacts and models to speed up repeated runs

[vecnode]

Implement a small caching layer for compiled kernels, downloaded model weights, and any expensive preprocessing artifacts, keyed by configuration and model version, to reduce wall‑clock time for repeated runs.

Why

• Many workflows (especially templates and CI) repeat similar runs with only minor config changes.
• Recomputing or re-downloading heavy artifacts each time wastes GPU and wall-clock time.
• A lightweight cache improves iteration speed without changing any single-run semantics.

What to do

• Define a cache directory convention (e.g., ~/.cache/disco-diffusion or a configurable path) and basic versioning scheme.
• Cacheable items could include:
  • Downloaded model weights / checkpoints.
  • Compiled CUDA kernels or other JIT artifacts, where applicable.
  • Precomputed embeddings or other expensive preprocessing outputs that are safe to reuse.
• Add hooks in the existing run path to:
  • Check for cached assets before recomputing/downloading.
  • Populate the cache on first computation.
• Ensure there is a simple way to invalidate or clear the cache when upgrading models or code.

Acceptance criteria

• Repeated runs with the same (or compatible) configuration reuse cached artifacts and are measurably faster.
• Cache behavior is transparent to users for single runs; correctness is unchanged.
• There is a documented way to clear or disable the cache if needed (e.g., env var or CLI flag).