Parallel execution of graph nodes

[shaheerzaman]

This PR introduces native support for parallel execution of graph branches and adds a convenience method for explicitly defining parallel successors.

Parallel branch execution:
The Graph now detects when a node has more than one outgoing edge and will run each branch concurrently via asyncio.gather, merging their outputs back into a single state. Nested parallelism is supported, and existing router behaviour (choosing a single successor via the "choice" key) remains unchanged.

New helper – add_parallel()
To make it explicit when multiple successors should be treated as parallel, Graph.add_parallel(source, destinations) wraps multiple add_edge calls. When executing, any node with more than one successor will automatically be run in parallel.

Internal refactoring:

Added _get_next_nodes() to return all successors of a node.

Added _invoke_node() to handle callback start/stop and streaming logic in one place.

Added a recursive _execute_branch() that walks a branch, spawns concurrent tasks for parallel branches, and merges their results.

Updated execute() to delegate to _execute_branch() and to clean up callbacks correctly at the end.