refactor(mecatui): standardize local async request tokens

Problem

The TUI currently uses a mix of Generation, Epoch, Seq, and action-specific names for local asynchronous request correlation. The values are used primarily as opaque request/response tokens, but some names imply ordering or time semantics and the current surface-local increment plus Model-level max handoff is harder to reason about than necessary.

Goal

Standardize local TUI async request correlation around explicit request tokens.

Proposed direction

• Use requestToken as the public/internal operation-level name.
• Add one Model-lifetime allocator (for example, nextRequestToken() uint64) that survives modal/surface teardown.
• Each in-flight operation stores the token it expects: page, transcript, action, etc.
• Response handlers compare tokens for equality only.
• Remove per-surface counter initialization and Model max reconciliation once all operations mint tokens through the shared allocator.
• Keep server cursors, server revisions, and protocol-provided generations distinct from local request tokens.

This is a cleanup/refactor. It must preserve the stale-response guarantees added by the /sessions surface migration and should add focused regression coverage for close/reopen and concurrent request classes.