perf(light-base): pre-compile cached parachain runtime during DB load

[replghost]

Summary

On warm start with cached runtime code (#3214), the sequence is:

1. Load DB → decode runtime code bytes
2. Start services (NetworkService, SyncService, RuntimeService)
3. Reach parachain Phase 2 → compile cached WASM (~200-300ms)

Compilation is pure CPU — it doesn't need network state or any service. Start compiling the moment the DB is decoded, before services are created. By the time Phase 2 runs, the compiled VM is ready.

Removes ~200-300ms from the critical path on every warm start.

Depends on

• perf(sync-service): skip parachain runtime download on warm start #3214 (cache runtime code in DB)

[replghost]

Starting work on this issue.

[replghost]

PR opened: #3226

[replghost]

Closing PR #3226 — A/B benchmarks across Paseo, Polkadot, and Kusama Asset Hubs showed no measurable difference. Compilation (~200-300ms release) is already hidden behind the 4-6s relay chain fetch. Not worth the added complexity. If relay fetch latency drops significantly, a simpler reorder (compile before peer wait) would be sufficient without spawn_task/oneshot machinery.