perf: Batch proof tasks at worker level

[yongkangc]

Problem

When proof tasks queue up in the worker pool (workers busy), each worker processes tasks one at a time. Multiple tasks for the same account result in redundant trie traversals.

The workers are the bottleneck (proof computation is slow), so tasks pile up in the worker queue, not the message channel.

Current vs Proposed

CURRENT: No batching
──────────────────────────────────────────────
Worker receives Task1 → compute → receives Task2 → compute → receives Task3 → compute
= 3 proof computations

PROPOSED: Drain & batch at worker
──────────────────────────────────────────────
Worker receives Task1 → drain Task2, Task3 → merge → compute once
= 1 proof computation

Solution

Apply the sparse trie drain pattern (sparse_trie.rs#L81-L94) to proof workers.

Batch here, where workers receive from the queue:

• Storage worker: proof_task.rs#L748-L749
• Account worker: proof_task.rs#L1028-L1029

// Current: process one at a time
while let Ok(job) = work_rx.recv() {
    process(job);
}

// Proposed: drain & batch
while let Ok(first_job) = work_rx.recv() {
    let mut batch = vec![first_job];
    while let Ok(next) = work_rx.try_recv() {
        batch.push(next);  // Drain pending
    }
    process_batch(batch);  // Merge & compute once
}

Note

This is complementary to #19168 (funnel-level batching). Funnel batching reduces tasks entering the queue; worker batching reduces tasks already queued.

Impact

• Fewer proof computations when workers are backlogged
• Reduced redundant trie traversals for same account

Related

• perf: Batch multiproof tasks at MultiProofTask message loop  #19168 (funnel-level batching)
• Tracking: batch multiproof tasks #19166 (tracking: batch multiproof tasks)