Add kernel runtime loop

[JaimeStill]

Summary

Implements the kernel runtime loop — the observe/think/act/repeat agentic cycle composing agent, tools, session, and memory. Config-driven initialization via New(*Config, ...Option) following the cold start pattern. Run() executes the loop until the agent produces a final response or iterations are exhausted.

Closes #14

Changes

kernel package (new)

• Kernel struct with New() and Run() implementing the agentic loop
• Config, DefaultConfig, Merge, LoadConfig for config-driven initialization
• ToolExecutor interface abstracting the global tools registry for testability
• Functional options (WithAgent, WithSession, WithToolExecutor, WithMemoryStore)
• ErrMaxIterations sentinel error

Agent interface evolution

• Conversation methods (Chat, ChatStream, Vision, VisionStream, Tools) accept []protocol.Message instead of prompt string
• Enables multi-turn conversations where the kernel passes full session history

ToolCall consolidation

• response.ToolCall eliminated in favor of protocol.ToolCall as single canonical type
• Custom UnmarshalJSON transparently handles nested LLM API format (function.name) and flat kernel format (name)

Subsystem configs

• session.Config with DefaultConfig, Merge, New (extension point for future backends)
• memory.Config with DefaultConfig, Merge, NewStore (nil store when path empty)

Convenience

• protocol.InitMessages(role, content) replaces verbose []protocol.Message{protocol.NewMessage(role, content)} across all call sites

Tests

• 14 kernel tests (93.8% coverage): direct response, tool calls, errors, max iterations, context cancellation, memory injection
• 6 ToolCall UnmarshalJSON tests (nested, flat, invalid, empty, array)
• Config and subsystem config tests