Parachain warm-start: reuse verified runtime from saved databaseContent

[replghost]

Summary

When a host saves and restores databaseContent (via chainHead_unstable_finalizedDatabase), the parachain runtime should be reusable without re-downloading the ~2 MiB :code from a peer. Currently, every parachain restart re-downloads the runtime even when the saved database contains valid runtime bytes.

Design

1. Persist the finalized parachain runtime snapshot in the database: :code bytes, merkle value, closest ancestor, bound to (blockHash, stateRoot).

2. On restore, verify the saved runtime via a lightweight merkle proof (~500 bytes) instead of re-downloading :code (~2 MiB). Request a sibling key in the trie that shares the same ancestor as :code — the proof includes the ancestor's child merkle values but not the leaf data. Compare the proof's merkle value to the stored hint.

3. Fall back to full peer download if verification fails (runtime upgraded, data corrupted, block pruned).

4. Pruning-window fallback: if the restored block is too old for peers to serve proofs (3 consecutive failures), discard the restored state and fetch a fresh finalized head from the relay chain.

Expected improvement

	Cold start (no DB)	Warm start (with DB)
Before	6-21s (#3203 fixes this)	6-21s (no benefit from saved DB)
After	6-21s (unchanged)	~1-3s (skip runtime download)

Measured 22x speedup on Kusama, 20x on Paseo in prior benchmarking.

Dependencies

• fix(sync-service): pass bootstrapped runtime to parachain runtime service #3203 (parachain cold start fix) — refactors bootstrap_parachain_consensus to return the runtime. This PR builds on that.
• fix(sync): handle stale warp sync bookkeeping after transition to all-forks #3185 (warp-sync transition fix) — independent, but both needed for full correctness.

Prior work

The implementation existed as part of the original #3185 (17 commits), now split into A (#3185), B (#3203), and this issue (C). Implementation branch: restored-warmstart-performance. Needs rebasing onto current main after B lands (bitswap/statement store removals changed the API surface).

Review feedback to address

• @skunert: don't clone the multi-MB wasm blob in DatabaseContent — use a flag or reference instead.
• @skunert: verify via merkle ancestor, not full :code re-download — already implemented in branch.

Benchmarks

Available in smolbench:

• bench/parachain-phase-breakdown.mjs — per-phase timing
• bench/parachain-matrix.mjs — full network matrix

[replghost]

Starting work. Building on top of #3203 (parachain cold start fix). Will port the warm-start implementation from the restored-warmstart-performance branch, resolving conflicts with recent main changes (bitswap/statement store removals).