ndigits0z(n::BigInt, b): avoid expensive off-by-one check for b=2^k.

Fixes #8727, but it's kind of a bandaid since for non-power-of-two
bases ndigits0z is still slow due to GMP's inexcusable laziness. At
least for power-of-two bases, however, BigInt hashing is now about
3x faster and allocates no memory.

(cherry picked from commit 19eb2fa)

base/gmp.jl

@@ -450,7 +450,7 @@ end
 function ndigits0z(x::BigInt, b::Integer=10)
     # mpz_sizeinbase might return an answer 1 too big
     n = int(ccall((:__gmpz_sizeinbase,:libgmp), Culong, (Ptr{BigInt}, Int32), &x, b))
-    abs(x) < big(b)^(n-1) ? n-1 : n
+    ispow2(b) || abs(x) >= big(b)^(n-1) ? n : n-1
 end
 ndigits(x::BigInt, b::Integer=10) = x.size == 0 ? 1 : ndigits0z(x,b)