Stream per-address pattern code (lazy Code from templates through writers)

`create2_deploys`/`create_preimage_deploys` with `code_pattern: unique_jumpdest_pre_amsterdam` materialize one unique 24,576-byte runtime per derived contract inside `Expand`'s returned slice, and `generator.Config.materializePreAlloc` then retains every blob in `Config.GenesisCode` for the entire run — peak RSS grows ≈24.6 KB per derived contract (measured: 247.8 MB at 10k contracts, GC-bracketed heap probe; linear → ≈3.7 GB at `spec-repricing-min.yaml`'s 150k, ≈37 GB at 1.5M), while `Storage` already streams via `iter.Seq2`.

PR #76's interim guard (added during review hardening) only warns above 2 GiB estimated residency and hard-fails above 64 GiB. Actually bounding residency requires:

- a lazy `Code` source on `templates.PreAllocEntity` (e.g. `CodeFn func() []byte` or a code iterator alongside `Storage`),
- teaching `generator.Config.materializePreAlloc` not to pin the bytes in `GenesisCode`,
- updating every client writer (geth, besu, nethermind, reth) to emit code per-entity and release it.

**Acceptance:** building `examples/spec-repricing-min.yaml` shows peak RSS independent of `salt_count`; cross-client genesis roots byte-identical to today's; the 2 GiB specbuild warning removed or demoted.

Related: `TODO(template-aware-budget)` in `internal/specbuild/build.go` (per-template `ProjectCost`) — a sibling blind spot in the same seam, kept separate.

Start with Expand and generator.Config.materializePreAlloc, then trace templates.PreAllocEntity into the geth, besu, nethermind, and reth writers. Use examples/spec-repricing-min.yaml to measure peak RSS and compare cross-client genesis roots with current output. Done means residency is independent of salt_count, roots remain byte-identical, and the 2 GiB warning is removed or demoted.

Written by the indexing model from the issue text.