refactor(framework): decouple Manager from TronJsonRpcImpl

[0xbigapple]

What does this PR do?
close #6963
Removes the Manager → TronJsonRpcImpl reverse dependency left by #6732, where core-layer Manager holds the json-rpc filter consumer through a @Lazy injection.

• Adds a standalone FilterCapsuleQueue bean; Manager.postBlockFilter/postLogsFilter produce into it, so Manager no longer references anything under org.tron.core.services.jsonrpc (repo-wide framework/src/main is now @Lazy-free).
• Moves the consumer loop into TronJsonRpcImpl: started by @PostConstruct when isJsonRpcFilterEnabled(), single daemon thread, stopped by close(). The consumer now runs on a daemon thread (the old Manager loop was non-daemon).
• TronJsonRpcImpl constructor becomes (NodeInfoService, Wallet, Manager); setManager() is removed and the direct-construction test sites are migrated.
• close() is rewritten: an AtomicBoolean guard makes it idempotent and serves as a visibility-safe stop flag, the loop exits on interrupt, and the consumer is awaited before logsFilterPool shuts down so an in-flight capsule completes on graceful shutdown. If the closing thread is interrupted, close() still waits up to 2 seconds, uninterruptibly, for the consumer before shutting down the pool.
• ApplicationImpl.shutdown() now closes the consumer after producers stop and before dbManager.close(); the later Spring-destruction close() is a no-op.
• The instanceof dispatch logs a warning for unknown FilterTriggerCapsule subtypes instead of silently dropping them.

Runtime behavior of the filter API is unchanged: same unbounded queue, same discard-on-shutdown semantics, one consumer shared by the FullNode/solidity/PBFT json-rpc services.

Why are these changes required?

Follow-up agreed in the #6732 review (see discussion). FullNode sets allowCircularReferences(false); @Lazy, like ObjectProvider or a runtime getBean(), only makes Spring tolerate the cycle. The reverse core → API edge stays, obscuring the component graph, and later json-rpc cleanups keep copying the pattern. This PR removes the edge instead.

This PR has been tested by:

• Unit Tests
• Manual Testing
  • private chain: block production with registered filters; eth_getFilterChanges delivers; SIGTERM log order confirms consumer → logs-filter-pool → dbManager → context close, single shutdown episode, no errors
  • Nile and mainnet lite nodes: 5 restart cycles each with 100 filters created and polled per cycle (~2,500 queries/cycle, zero rpc errors); on mainnet the log path delivered 20k+ real contract log entries through the new queue; filters correctly report filter not found after restart; no RejectedExecutionException, no bean-cycle errors, shutdown order identical in every cycle

Follow up

Extra details

[bladehan1]

LGTM