use AutoAnnotationPipeline

[ArneBinder]

This makes all inference and evaluation workflows fully independent of PyTorch-IE (src/train.py is still PyTorch-IE specific). This prepares for PIE projects that do not rely on torch, but e.g. LLM workflows (see "Future" in ArneBinder/pie-core#17).

Changes:

• use pie-core 0.3.1 since this PR requires Auto.from_config: allow to pass the class name per keyword argument, even if no respective key is in the config pie-core#95
• switch back to use pie_core.AutoAnnotationPipeline instead of pytorch_ie.PyTorchIEPipeline
• add comments to add pipeline_type: pytorch-ie for old models to pipeline configs
• actually set pipeline.pipeline_type: pytorch-ie in train.yaml config
• save the pipeline after training (instead of individual model and taskmodule)
• during inference, raise an exception when trying to load a model checkpoint into a pipeline that is no PyTorchIEPipeline

cc @RainbowRivey