Improve visibility into long-running subagent activity

## Problem

When I hand off long reviews or implementation tasks to subagents in the Agents window, they can run for 15 minutes or more while the main session appears idle or stuck.

The current spinner does not help me tell whether a subagent is making progress, waiting for something, blocked, duplicating another worker, or stalled. This makes the parent session feel non-interactive and leaves me unsure whether I should keep waiting or intervene.

This is especially confusing when similarly named subagents are working on the same problem with different models.

## Suggested first step

Show a compact status for each running subagent with:

- The model being used
- The current activity or active tool
- Elapsed time

For example:

> Reviewing tests · GPT-5 · 12m 34s

Token, credit, or context usage could be available as secondary information when expanded, without making the default conversation view feel like a monitoring dashboard.

## Possible follow-ups

Other useful improvements could include:

- Clear states such as waiting, blocked, approval required, failed, cancelled, or completed
- An indication when no meaningful progress has happened for some time
- A parent-level overview of active subagents, especially when several are running in parallel
- Recent actions, changed files, or test status in expanded details
- Controls to open, steer, cancel, or retry a subagent
- A warning when multiple subagents appear to be doing duplicate work

## Expected outcome

Users should be able to quickly distinguish productive work from waiting or stalled work, understand which subagent is doing what, and decide whether intervention is needed without opening every child conversation.

cc @sandy081