pub fn cast_to_master(tensors: &[Tensor], master_dtype: DType) -> Vec<Tensor>
Cast a list of tensors back to master dtype for gradient accumulation.