Fix interleaved_numeric_conversion.h

[yz-chen18]

uint4 fast cast should minus 64.5 for elt_45 and elt_67, which is 1032 / 16

[byshiue]

Can you explain what problem do you encounter? How do reproduce the issue you encounter?

[rhenry-nv]

@yz-chen18 In addition to @byshiue's question, I am not sure your suggestion is correct. Due to the bit positioning, we have:

elt_45/elt_67 = 1024 + 16X

Then, we divide by 16 with fma to give us 64 + X. However the original value is biased to be unsigned so, X = O + 8 where O is the original, signed value. So to extract the final value we must do X - 64 - 8 which gives -72. Do you have an example where that produces incorrect results?