A divisor with an N-bit multiplier, which computes the quotient as MULUH(m, n') >> shift
for n' from $body. Up and Down differ only in n', and the macro keeps them from
duplicating the rest. |n| ... looks like a closure but is macro syntax for the expression,
because a const fn can’t call a closure
Whether a saturating increment on u32 lanes is cheap: one uqadd on aarch64, and pminud
from SSE4.1 and a subtract on x86. Baseline x86-64 has no unsigned 32-bit min and emulates it
with compares, so there the saturating increment of DivisorU32::Down costs more than the
add and halve of DivisorU32::AddIndicator