lite: promote precision of quantization multiplier

The original implementation loss precision when calculating
`input_product_scale`. This patch fixes it.

Test: bazel run tensorflow/lite/kernels:kernel_util_test