Follow up for #5519

1. Scope guard for multiple return. This pattern is used in other places where we have multiple return.
2. Drop _bittestandreset. The compiler emits btc for the ^= 1 << n pattern, which has the same performance as btr for _bittestandreset.
   • For search could also use _Bingo &= _Bingo - 1, or _blsr_u32, but there's not enough evidence in significant enough performance advantage to use this
   • For find_end could try _bzhi_u32, but didn't even bother, here the expectation is even lower.
3. Always use unsigned 32-bit for bit masks from movemask*. Always 32 bit because internally there is 32-bit destination register, always unsigned because bitmask.
   • Still int for cmpestr* masks, it is compile time, and only few bits, and intrinsic takes int
   • In some places the cast to other bit width is needed, it is kept there