Add a KV cache to nanoGPT generation and prove equivalence

Cache keys and values per layer, handle positions and masks for incremental decoding, and test that cached logits equal full-forward logits at every step.

The exercise happens on Self-directed project (nanoGPT as the base). Canonical link: https://github.com/karpathy/nanoGPT. Link check: link responded (5 Oct 2026). Opened from this app 0 times; opening is not counted as an attempt.

Supports: Inference-path understanding for research-engineering and AI-infrastructure roles.