feat: add multi-model support and environment variable API keys

[toadlyBroodle]

Description

This PR adds support for using different LLM models for planning vs code implementation phases, along with environment variable support for API keys. It also refactors duplicated code into shared utility functions.

Related Issues

• Addresses the need for flexible model configuration (use powerful models for planning, faster models for coding)
• Improves security by supporting environment variables over hardcoded secrets

Changes Made

Features

• Add separate planning_model and implementation_model config options for all providers (Google, Anthropic, OpenAI)
• Support GOOGLE_API_KEY, ANTHROPIC_API_KEY, OPENAI_API_KEY environment variables
• Environment variables take precedence over mcp_agent.secrets.yaml values
• Add .env.example template for easy setup

Refactoring

• Add shared get_api_keys() and load_api_config() functions in utils/llm_utils.py
• DRY up duplicated config loading code across workflow files (~40 lines removed)
• Use lazy imports for anthropic and openai packages to make them optional dependencies
• Remove hardcoded brave search server references, use config-based search_server_names

Config Changes

• Add planning_model and implementation_model options for google, anthropic, and openai sections
• Add mcp_agent.secrets.yaml to .gitignore (users should use .env instead)

Checklist

• Changes tested locally
• Code reviewed
• Documentation updated (if necessary)
• Unit tests added (if applicable)

Additional Notes

Breaking Changes: None. All changes are backward compatible - existing configurations continue to work.

Migration: Users can optionally:

1. Copy .env.example to .env and add their API keys there
2. Add planning_model and implementation_model to their config if they want different models for each phase

Example config:
google:
default_model: "gemini-3-pro-preview"
planning_model: "gemini-3-pro-preview" # Used for analysis
implementation_model: "gemini-2.5-flash" # Used for code generation