mat/big: more optimal Karatsuba threshold

benchmark           old ns/op    new ns/op    delta
BenchmarkHilbert      6253043      6267289   +0.23%
BenchmarkMul         45355940     39490633  -12.93%

R=r
CC=golang-dev
https://golang.org/cl/6355104

src/pkg/math/big/nat.go

@@ -236,7 +236,7 @@ func karatsubaSub(z, x nat, n int) {
 // Operands that are shorter than karatsubaThreshold are multiplied using
 // "grade school" multiplication; for longer operands the Karatsuba algorithm
 // is used.
-var karatsubaThreshold int = 32 // computed by calibrate.go
+var karatsubaThreshold int = 40 // computed by calibrate.go
 
 // karatsuba multiplies x and y and leaves the result in z.
 // Both x and y must have the same length n and n must be a