v0.14.1

• Execution stages are now a class-level declaration. Node gains
  EXECUTION_STAGES: frozenset[ExecutionStage] (default {ALWAYS}) beside _category /
  _tags; a subclass declares its stages on the class body (a set literal or stage names are
  normalized to a frozenset at class creation, invalid entries raise at import) and no longer
  needs to re-plumb an execution_stages default through every __init__. The constructor
  argument stays as the per-instance override, which is what a pipeline yaml's
  hparams: {execution_stages: [inference]} reaches through PipelineFactory; string values
  are coerced and an unknown name ([Inference], [infer]) raises ValueError naming the node
  instead of silently never running it. node.execution_stages is a read-only property
  returning the frozenset; assigning to it raises AttributeError unless the subclass opts in
  with EXECUTION_STAGES_MUTABLE = True. Node.get_execution_stages() mirrors
  get_category() / get_tags(). consume_base_kwargs is unchanged and still works.
  Breaking for code that assigned node.execution_stages = {...} after construction
  (two example scripts in cuvis-ai-unet do): pass the stages to the constructor, or declare the
  class mutable. should_execute, serialization (stages are still not part of hparams) and the
  pipeline's stage pruning are unchanged.

What's Changed

• Declare execution stages on the node class by @nghorbani in #70

Full Changelog: v0.14.0...v0.14.1