libc: improve accuracy of log1pf. Fix #515

Improved log1p version, using Pade approximant,
4 multiplications and one division, relative error is bounded
by about 2^-21=5e-7.

From #516 (comment)

Author: parisseb

src/libc/log1p.c

@@ -1,7 +1,13 @@
 #include <math.h>
 
-float log1pf(float x)
-{
+float log1pf(float x) {
+    if (fabs(x) <= 0.125) {
+        // pade(series(ln(1+x),x=0,6,polynom),x,5,3)
+        // (-57*x**2-90*x)/(x**3-21*x**2-102*x-90)
+        // relative error less than 1e-7
+        return x*(57*x+90)/(((21-x)*x+102)*x+90);
+    }
+    // relative error about 2^-21 if abs(x) is just above 0.125
     return logf(1 + x);
 }