Interface ConvertOptions

Replace each quantized module with its dense weight and drop the config's quantization block.

Cast every floating tensor (quantization scales/biases included) to this dtype.