feat: Implement a memoization-based cache

[matheus23]

This is a performance optimization for CAR mirror.

This helps, especially when DAGs are huge, blockstores are locally slow (e.g. reading from disk) and when latency is not the main issue anymore.

It improves benchmarks quite a bit:

---

What this PR is doing is simply implementing one version of #28: A cache for block references: Cid -> Vec<Cid>.
This way, when there's a cache hit, we don't even need to fetch the block, parse it and find links. Instead, we get further references directly.

In some cases, this means running the same operation twice won't fetch any block from the blockstore at all.

Since different environments will want to implement different types of caches (e.g. the quick-cache based cache won't work in Wasm), the cache implementation is abstracted behind a trait Cache.

I've bundled two implementations: NoCache that simply always re-computed the value (no memoization), and InMemoryCache which uses the quick_cache library.

---

This PR depends on #29

[matheus23]

Closes #28