Replaces the process-global LRU with a chain-owned, byte-bounded
SizeConstrainedCache (128MB) shared between block processor and
prefetcher. Threaded through Process/Prefetch as an optional argument;
RPC, tracing, and stateless callers pass nil and the EVM falls back to
its per-instance map. Adds chain/cache/jumpdest/{hit,miss} meters.
