feat[bc]: rename gauges, add seeder metrics, and eagerly open blob core on indexers

[yuranich]

What problem does this PR solve?

Four observability gaps that surfaced on the first staging scrapes after #1689 landed:

1. _total suffix on gauges violates Prometheus / OpenMetrics naming conventions. qvac_registry_models_total and qvac_registry_blob_cores_total are gauges (values go up and down). Scrape linters flag them and OpenMetrics parsers can reject them.
2. RPC counters export no series before the first RPC call, so rate() returns NaN on fresh dashboards — panels look broken on cold start.
3. qvac_registry_blob_core_* metrics are empty on indexer nodes that don't use QVAC_BLIND_PEER_KEYS mirror replication — the blob core is populated into blobsCores lazily (only on addModel or _setupBlindPeering), so on an indexer that runs the topic-pull flow it stays Map(0) indefinitely.
4. No visibility into replication durability. Operators have no way to see how many full replicas of the view core or any blob core exist in the swarm. hypercore_total_peers is too coarse (unions all unique UDX streams across every hypercore, including RPC-only clients).

How does it solve it?

• Renames qvac_registry_models_total → qvac_registry_model_count and qvac_registry_blob_cores_total → qvac_registry_blob_core_count. RPC counters keep _total — they're actual counters.
• Pre-initialises both RPC counters with the five known methods (add-model, put-license, update-model-metadata, delete-model, ping) at 0 in the QvacMetrics constructor, so rate() returns 0 from the first scrape.
• Eagerly opens the blob core at the end of _open() when base.isIndexer || base.localWriter, so blob_core_peers, blob_core_byte_length, blob_core_fully_downloaded, etc. populate on indexers that use the topic-pull blind-peer flow. Reader-only nodes skip this (their writable: true would create a local core with the wrong key).
• Adds replication-durability gauges:
  • qvac_registry_view_core_seeders — peers that hold the view core fully and advertise remoteUploading. View is small (a few MB of autobase metadata), so this converges to connected-peer count within an RTT — no separate raw-peers metric needed.
  • qvac_registry_blob_core_seeders{core_name} — same signal per blob core. Paired with the existing qvac_registry_blob_core_peers, the peers - seeders gap exposes peers currently downloading vs. serving.
• Updates DEPLOYMENT_GUIDE.md — metric table, replication-durability alerting guidance, note that _seeders metrics use p.remoteOpened && p.remoteUploading && p.remoteContiguousLength >= core.length so partial/mid-handshake peers aren't counted.
• Updates the in-tree Grafana dashboard — two panels retargeted at the new metric names.
• Extends the metrics integration test — asserts new names exist, legacy names are gone, RPC counter series are pre-initialised to 0, and view_core_seeders is exported as a single series that is 0 with no connected peers.

Verified with npm run lint, npm run test:unit (37/37), and npm run test:integration (30/30, 146/146 asserts).

Breaking changes

• qvac_registry_models_total → qvac_registry_model_count
• qvac_registry_blob_cores_total → qvac_registry_blob_core_count

Anyone scraping these series must update dashboards / alerts. The in-tree Grafana dashboard is updated in this PR; no other consumers are known.