Support per-agent extended thinking configuration in subagent frontmatter

When using custom subagents (.claude/agents/*.md), there's no way to control extended thinking per-agent. The only options are global (alwaysThinkingEnabled in settings.json, session toggle, MAX_THINKING_TOKENS env var).

Problem: Users who run diverse agent types — from simple mechanical agents (calendar invites, template-based documentation) to complex reasoning agents (deep research, validation, planning) — are forced into all-or-nothing. Either all agents use thinking tokens or none do.

Proposed solution: Add a thinking field to the agent .md frontmatter:

# .claude/agents/calendar-invite.md
---
name: calendar-invite
description: Send Google Calendar event invites
tools: Bash, Read
model: haiku
thinking: false
---

# .claude/agents/validator.md
---
name: validator
description: Critically validate plans and code
tools: Glob, Grep, Read, Write
model: sonnet
thinking: true
thinkingBudget: 10000
---

This would allow users to:

• Disable thinking for mechanical agents to save quota
• Enable thinking for reasoning-heavy agents where it matters
• Optionally cap thinking budget per-agent

Especially important for quota management - thinking tokens on opus are the most expensive contributor to the 5-hour rolling window, and there's no way to be selective.