WS fixes + improvements

[stefangutica]

Reasoning

Custom WebSocket subscriptions queried Elastic three times per round — once per gateway — even though the transactions query hits the same operations index as the transfers one, so the same data was fetched and enriched twice on the latency-critical path.

Proposed Changes

• Add CustomSubscriptionsDataFetcher: fetches a round's data once and derives the transactions payload from the transfers result. The three custom gateways no longer query Elastic — they receive the data and only match rooms and emit, so their push methods are now synchronous.
• Drop canBeIgnored operations from custom transfers, via a new internal withCanBeIgnoredFlag query option so the flag survives entity mapping into Transaction.
• Read the latest round timestamp straight from the rounds index instead of going through RoundService (which also ran a getCurrentEpoch query first), and drop RoundModule from the subscription module.
• Start getStats in parallel with the round lookup, and persist the round cursor inside the loop so a polling timeout no longer re-broadcasts rounds that were already sent.
• NetworkGateway.pushStats calls getStats(true) to bypass the cache.
• broadcastIntervalMs 1000 → 600 on devnet and testnet.

How to test

• Subscribe to custom transactions/transfers/events and confirm the same messages arrive as before, with refunds no longer delivered on the transfers channel.
• Check the Elastic query count per round drops from three to two.