feat(prover-node): tear down checkpoint sub-tree once block proofs are ready (A-1213)

[PhilWindle]

What

Reduces prover-node memory by tearing down each checkpoint's sub-tree orchestrator as soon as its block proofs are captured, instead of retaining it for the whole proof-submission window.

Each CheckpointProver held its CheckpointSubTreeOrchestrator — and every TxProvingState, AVM circuit input, and base/merge/parity proof it carried — alive until the prover was reaped at the end of the submission window, long after the checkpoint's block proofs were produced. Across every proven checkpoint this accumulates and dominates prover-node memory under load.

Now, once the sub-tree's SubTreeResult is captured, the CheckpointProver drops the sub-tree orchestrator. The block-proof outputs survive via the resolved promise; post-completion consumers (the top-tree job, a rebuilt EpochSession, failure upload) read only whenBlockProofsReady() and the prover's own fields (checkpoint, txs, headers, sibling paths), never the sub-tree. Teardown is tracked so whenDone() awaits it and stays idempotent with the cancel/reap path.

Correctness

• Teardown of the sub-tree runs exactly once: teardownSubTree() captures the orchestrator and nulls the field in a synchronous prefix before any await, so the success-driven teardown and a concurrent cancel/reap can't both call stop().
• whenDone() awaits the block-proof settlement between runPromise and the teardown check, so it can't report completion in the window where block-level proving has finished enqueueing (runPromise resolved) but the sub-tree proofs — and the teardown they trigger — are still outstanding.

Scope

This reclaims the cross-checkpoint retention of already-proven sub-trees. It does not free the intermediates of checkpoints that are still proving (the in-flight AVM inputs / base-rollup hints / recursive proofs). Freeing those as they are consumed is a separate, more invasive change and is deliberately left as a follow-up.

The broker remains a separate large consumer (A-1215), as does the prover-node's BrokerCircuitProverFacade inline-inputs retention (A-1517).

Testing

• yarn build, full @aztec/prover-node checkpoint-prover suite.
• New tests: the sub-tree is released exactly once on completion while the block-proof outputs still resolve; a subsequent reap cancel is a no-op; and whenDone() stays pending until the sub-tree result lands and teardown completes (asserted by calling it before the result resolves).