parallel exponent chunks to regain multiexp performance regression

[XuyangSong]

If we parallel the exponent into chunks, we can regain some performance regression.

multiexp/65536          time:   [409.96 ms 411.86 ms 413.81 ms]                           
                        change: [-8.6071% -7.8501% -7.1148%] (p = 0.00 < 0.05)
                        Performance has improved.