Multi-LLM routing by task type (hybrid model setup)

[0xallam]

Add support for using multiple LLMs in the same run, automatically selecting the best model per task/sub-agent. Example: fast/cheap models for simple steps and background subagents, stronger reasoning models for complex exploitation/debugging, and vision-capable models for browser-heavy workflows.