fix: make pipeline step settings optional and fix TrOCR device mismatch

[andy778]

Description

pipeline/steps.py

• Made settings field in PipelineStepConfig optional (dict | None = None), so pipeline YAML configs no longer require an explicit settings: key for steps that use defaults.
• Added a or {} fallback in init_step so a None settings value is safely passed as an empty dict to from_config.
  models/huggingface/trocr.py
• Added a workaround for a device mismatch bug in transformers' TrOCRSinusoidalPositionalEmbedding. During inference, position_ids are on CUDA while self.weights remains on CPU (or stored as a meta tensor from the checkpoint). This causes a runtime error.
• Fix: after loading the model, recompute the sinusoidal positional embedding weights from scratch using the layer's own get_embedding() method and move them to the correct device.

Why
These two fixes address runtime errors encountered when running TrOCR-based pipelines:

1. A ValidationError when a pipeline step config omits the settings key.
2. A device mismatch error (Expected all tensors to be on the same device) during TrOCR inference on CUDA.

Issue ticket number and link from

#47

Checklist

• I have performed a self-review of my code locally
• [x ] My code follows the style guidelines of this project (e.g. passes ruff lint + mypy type hinting)
• If it is a core feature, I have added thorough tests.
• [x ] Documentation has been updated or the changes are too minor to be documented