feat(tui): phase labels for non-bootstrap phases (LLM call, schema diff, review lanes)

Deferred from the v0.9.3 pre-release review.

### Current behavior

`packages/tui/src/util/phase-label.ts` maps five `bootstrap.*` phase names to human labels ("Loading session...", "Discovering tools...", etc.). Once bootstrap completes, every subsequent phase name falls through to "Thinking..." — the same default the idle spinner uses.

### Gap

For long-running dbt review runs (schema diffing, LLM lane, tool listing), the label stays on "Thinking..." for the whole run — no signal that the tool is actively doing work versus stalled. This reduces the value of the bootstrap phase-label UX added in this release.

### Proposal

Extend the phase-label map with per-workload labels (e.g. `review.diff`, `review.equivalence`, `review.ai-lane`) once the latency-instrumentation from the bootstrap trace spans provides enough signal to pick meaningful groupings. Start with the review pipeline since it has the highest wall-clock and the loudest complaint surface.

### Why deferred

The bootstrap instrumentation just landed. Picking non-bootstrap phase groupings without the latency data risks labels that don't match the actual work — a follow-up when we can point at real spans.

Start with packages/tui/src/util/phase-label.ts to understand the existing bootstrap mappings and fallback behavior. Then inspect the review pipeline and the latency-instrumentation trace spans to identify meaningful non-bootstrap groupings. Done means the selected review phases have accurate labels grounded in observed spans and the labels no longer use the generic idle wording during those workloads.

由索引模型根據 Issue 內容生成。