Vectorized infer_beam_batch for improved performance

[9173860]

This pull request introduces key optimizations and improvements in model_48px aimed at increasing the efficiency of the OCR pipeline:

Vectorized infer_beam_batch:

The infer_beam_batch function in model_48px has been vectorized, significantly improving the speed of OCR inference. The original infer_beam_batch is also kept, so you can easily switch between the two for testing purposes.

Refactored encoders and decoders with forward methods:

Extracted the forward methods for both encoders and decoders. This refactoring enables more straightforward model exporting (e.g., to ONNX), allowing further optimization of inference performance and integration with deployment platforms like Triton Inference Server.