Fix CellMap training issues: resolve NaN losses and infinite epochs

[BoyuShen2004]

• Add data type conversion (_prepare_images/_prepare_labels) to prevent uint8/float16 mismatch
• Implement CellMapBalancedLoss with comprehensive NaN/inf handling for sparse segmentation
• Add batch skipping for empty batches to prevent NaN contamination in progress bars
• Replace problematic pos_weight computation with uniform weights to avoid numerical instability
• Add limit_train_batches/limit_val_batches to prevent infinite epochs from large datasets
• Enhance loss function with gradient-connected safety values and extensive numerical stability checks