roundup_power_of_two performance

[SiarheiVolkau]

It seems that roundup_power_of_two is over-engineered.

Boiling down such thing leads to typical count leading zeros (clz) problem.
Modern CPUs will do it in a few cycles via special instruction(s),
we just need to use __builtin_clz for suitable compiler or rewrite it
in a way that allow compilers detect clz pattern in the code.

In current implementation no one compiler detects clz pattern thus it is just slow.