Reconsider parameters used for delay model regression curve fit for multiplies (umul, smul)

Currently, the sky130 delay model uses just the bit width of the first operand when estimating a particular multiplication's hardware delay.

Other delay models use just the *result* bit width for mult delay estimation.

Neither is sufficient since they are independent and both contributing factors. In fact the second operand's bit width also contributes, although perhaps the two operand widths can be combined into a single expression, so that the regression curve fitting has two expressions (combination of operand widths, and result width).

Consider:
u8 * u8 --> u8
u8 * u8 --> u16
u8 * u16 --> u16