Reduce embedded IPFS node resource usage

[aopoltorzhicky]

Profiling a metadata-indexer instance showed that ~95% of memory (~2GB RSS) and CPU
was consumed by the embedded IPFS node acting as a full network participant:
announcing every cached block to the DHT and serving blocks to other peers.

Changes

• Apply config to existing repositories. fsrepo.Init is a no-op when the repo
  already exists, so config changes in code never reached deployed nodes. The config
  is now persisted on every start via repo.SetConfig (identity is preserved).
• Disable providing and bitswap server (Provide.Enabled=false,
  Bitswap.ServerEnabled=false) — the node is a pure consumer.
• Switch to delegated routing (Routing.Type=delegated) — provider lookups go
  through HTTP routers instead of running a DHT client.
• Lower connection manager watermarks to 20/50 with a 1m grace period.
• Bound the datastore: StorageMax=2GB + a periodic GC goroutine
  (corerepo.PeriodicGC); the enable-gc ExtraOpt is daemon-only and had no effect.
• Skip repo re-initialization on start when it already exists (avoids generating
  and discarding an RSA key on every launch).
• Remove the reconnect goroutine that logged the full peer list every 3 minutes.

Tests

Added tests covering repo creation (idempotency, identity preservation) and config
persistence/override on existing repos — no node startup or network required.