v0.15.0

• Execution stages are class-attribute only (breaking). Node.__init__ no longer takes
  execution_stages=: the stages a node runs in are exactly its class's EXECUTION_STAGES.
  Gone with it are the per-instance state, the execution_stages setter and the
  EXECUTION_STAGES_MUTABLE opt-in (node.execution_stages is a read-only property returning
  the class declaration; assigning to it raises AttributeError), and Node.consume_base_kwargs
  (pass **kwargs straight to super().__init__). A pipeline yaml may still carry the inert
  hparams: {execution_stages: null} that earlier releases wrote; it is accepted and dropped.
  Any other value (execution_stages: [inference]) fails at load with a TypeError naming the
  node: picking a different class is the only way to change a node's stages, so running the
  TensorBoard monitor at inference is no longer a yaml opt-in (Predictor.predict(stage= ExecutionStage.TEST) runs the val/test nodes over a dataset instead). Plugins that already
  declare their stages on the class (adaclip 0.3.2, dinomaly 0.6.4, inspecscrap 0.3.0,
  unet 0.3.0, augment 0.4.3) need no change; a subclass that still names execution_stages
  in its own __init__ signature must drop it.

What's Changed

• Make execution stages class-attribute only (0.15.0) by @nghorbani in #71

Full Changelog: v0.14.1...v0.15.0