docs: add Agent Instruction Anti-Patterns guide

[jpantsjoha]

Summary

Add comprehensive documentation identifying common agent instruction anti-patterns and providing proven solutions. This guide addresses community-reported issues with HITL (Human-in-the-Loop) patterns and autonomous polling instructions.

Problem Addressed

Based on analysis of GitHub issues #3607, #3645, and #3184, developers are struggling with:

• Agent instructions that create polling loops, increasing inference costs by ~93%
• Poor user experience from manual "continue" clicking in HITL scenarios
• Confusion about when to use different HITL patterns
• Lack of guidance on proper agent instruction architecture

Key Features

• Identifies autonomous polling anti-patterns that violate LLM stateless principles
• Documents Backend Blocking Poll pattern as the proven solution used in ADK framework
• Provides decision matrix for choosing appropriate HITL approaches
• Includes validated code examples cross-referenced with ADK codebase samples
• Covers instruction overloading and separation of concerns best practices

Validation

All patterns and solutions have been validated against:

• /contributing/samples/human_in_loop/ - Complete HITL example implementation
• /contributing/samples/human_tool_confirmation/ - Tool confirmation patterns
• /src/google/adk/tools/long_running_tool.py - Framework-level anti-polling protection
• /src/google/adk/flows/llm_flows/functions.py - Async execution patterns

Impact

This documentation fills a critical gap identified in community discussions, providing:

• Cost reduction by preventing inefficient polling patterns
• Better UX through proper async HITL implementation
• Clear guidance for developers on instruction best practices
• Framework alignment with ADK's event-driven architecture

Testing

• ✅ Markdown syntax validated
• ✅ Navigation structure updated in mkdocs.yml
• ✅ Examples cross-referenced with working ADK samples
• ✅ Links to GitHub issues verified

The guide provides immediately actionable guidance for ADK developers building production agent systems.