Investigate and improve interrupt flag codegen

[bugadani]

Based on #2960 we might have some work to do - for one, we should change the other drivers as well. But we could also look into why EnumSet doesn't optimize well, too.

Ideally, if we could coerce the interrupt flag enums to use discriminants equal to their corresponding bit values (which only works if the flags are in a single register, and may be device dependent), theoretically the codegen could just be as little as read-and-mask. Can we somehow achieve that?

[wisp3rwind]

I had a quick look at EnumSet when creating that PR.

If I understood it correctly, the way that the iterator works is as follows:

• create a copy of the data (typically some unsigned integer)
• in each step of the iteration, count trailing zeros to find the first set bit; then toggle that bit and yield the index

I'd guess that the compiler gives up on constant propagation through the second step and decides not to unroll the loop?