Add deeplearning reconstruction metrics

[sergioald]

This PR adds reusable reconstruction metrics for the deeplearning module.

The metrics are implemented in a separate metrics.py module so they can be reused beyond the autoencoder classes and also used as differentiable PyTorch losses during training.

Changes include:

• Add reconstruction_error()
• Add evaluate_reconstruction()
• Add ReconstructionLoss
• Support MSE, MAE and RMSE
• Support none, sample, mean and sum reductions
• Support both NumPy arrays and PyTorch tensors
• Preserve gradients when PyTorch tensors are used
• Add tests for NumPy evaluation, PyTorch differentiability, mixed NumPy/tensor inputs, loss usage and input validation

This PR intentionally does not modify the autoencoder classes yet. A follow-up PR can add convenience methods to the autoencoder API that call these shared metrics internally.