feat(cli): add /usage for per-turn token timeline

[emal-avala]

Summary

Adds `/usage` — a per-turn token timeline table.

#	model	input	output	cache read	cache write
1	claude-opus-4-7	1204	430	0	8100
2	claude-opus-4-7	320	210	8100	0
∑		1524	640	8100	8100

Followed by a one-line cache hit rate hint. `/cost` stays the go-to for aggregate cost; `/usage` is for "where did my tokens go this session?".

Why

`/cost` already shows per-model aggregates, but you can't see a timeline — was turn 3 the one that ate the context, or turn 7? The per-turn table answers that without needing to dig through tracing logs.

Handles mixed-model sessions: each row picks the model that turn actually used (the `model` field on the assistant message), falling back to the session's current model when it's missing.

Implementation

• `collect_usage_rows` walks `state.messages`, picking `Message::Assistant` entries with a `Usage` payload. Turn index = 1-based counter over counted turns.
• Column widths computed from the data (model column capped at 24 chars; long names truncated to the tail to keep family tokens visible).
• Cache hit% = `cache_read / (input + cache_read + cache_write)` across the session.

Test plan

• `cargo fmt --all` clean
• `cargo check` passes
• `cargo clippy --no-deps` clean
• `cargo test --bin agent commands::tests::usage_rows_skip_messages_without_usage` passes (verifies rows skip assistant messages without usage and fall back to default model)
• Manual: multi-turn session → `/usage` prints table with a row per turn
• Manual: fresh session → `/usage` prints "No completed turns with usage data yet."

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.