Increase exponential functions precision

[apelloni]

To keep track of the things being fixed for the exponential functions and increase the precision.

exp() was precise only up to f64 and with #42 matches TwoFloat.

TODO:

• exp_m1() has similar issues where now a.exp() -1.0 is more precise than running a.exp_m1()
• a.exp2() less precise than running (a*LN_2).exp()
• adjust tests to reflect a higher precision as expected from double-double arithmetic

Unrelated to EXP but exposed while improving the precision

• log10 poor precision due to 1/LN_10 lost accuracy
• Fix division
  • f64/ TF
  • TF/TF
  • TF/f64?
  • TF/=TF
  • TF/=f64?

[apelloni]

both exp_m1 and exp2 where affected by wrong coefficients in the various EXP#_COEFFS arrays.
Now, they all use the same FRAC_FACT array containing the inverse factorial coefficients.

I've updated them in fbdc57c and 2958f6d.

I use the rescaling trick also for exp2 to improve the convergence, where 2^r is rewritten as (2^r')^512