[stmartineau99]

Changes

• dinov3 backbone: _get_checkpoint loads the encoder via microsam2 get_vfm_model, like dinov2
• this means that training with ViT backbones is dependent on having microsam's dev branch installed for now
• bf16 inference: new dtype arg in get_prediction, torch.bfloat16 runs the model under autocast

Tests

• added test_prediction_bfloat16 test to compare bfloat16 autocast predictions with default float32