Since this will only ever be done at compile time (as guaranteed by
using `consteval`), we can afford to prioritize precision over speed.
To compute an Nth root, we simply do a binary search over representable
floating point numbers, looking for the number whose Nth power most
closely matches the original number.
Fixes#494. We have included a test case reproducing the original
problem exactly. All tests use "within 4 ULPs" as the criterion, which
is (I believe) equivalent to the googletest `EXPECT_DOUBLE_EQ`
criterion.
Chip Hogg
7e894788d7