feat: support parallel tool calling

[edcuba]

Add support for executing batches of tool calls in parallel.

Changes

• Added the parallel_tool_calls option to LlmModel and supported model implementations.
• Enabled parallel tool execution by default.
• Execute independent, non-confirmation server tools concurrently with asyncio.gather.
• Continue handling client tools, confirmation-required tools, and mixed tool batches sequentially.
• Preserve interrupted tool requests when authentication yields during parallel execution, while avoiding duplicate execution of completed tools.
• Added serialization support for the new model option.
• Documented parallel tool calling, including guidance for tools with side effects or shared state.
• Updated the changelog.

Tests

Added integration coverage for:

• Concurrent execution of parallel server tools.
• Disabling parallel tool execution.
• Resuming a parallel batch after an authentication interruption.
• Ensuring completed tools are not executed twice after resumption.