[Feature] Ring-buffer backing store for Mailbox to remove O(n) shift()

[pathosDev]

Rationale

Mailbox user/system queues are plain arrays and dequeue via Array.shift() (src/internal/Mailbox.ts:51, 55) — O(n) at queue depth. At the default BoundedMailbox capacity of 10k a deep mailbox pays a substantial reindexing tax per message. prependUser uses spread-unshift (O(n+m), stack risk for very large batches), and the ThroughputDispatcher work queue also shift()s (src/Dispatcher.ts:82).

Scope

• Circular head/tail ring buffer in the base Mailbox so BoundedMailbox inherits it (the default mailbox is the one that must get faster).
• Amortized O(1) enqueue/dequeue; bulk prependUser without spread.
• Same treatment for the ThroughputDispatcher queue.
• New deep-queue benchmark case alongside the existing single-node benches.

Documentation

• CHANGELOG entry (internal performance; no user-facing API change expected).

Acceptance

• fast-check model test: the ring behaves identically to a reference array model under random enqueue/dequeue/prepend/suspend sequences (this also guards against regressions of [Bug] BoundedMailbox drop-head overflow is a no-op while the mailbox is suspended #407).
• Deep-queue bench (depth ≥ 1k) strictly faster than baseline; shallow-queue benches (tell-throughput, priority-mailbox) show no regression.

Relates

Improvement program M2. #27 (benchmarks), #407.