Improve performance of x**y when y is a huge value

[tompng]

When y.exponent is several thousand or more, x**y was slow because exponentiation by squaring requires several thousands of multiplications.
Use exp and log in such case.

irb(main):002> n=BigDecimal('1e+4000'); (1+1/n).power(n, 100)
processing time: 14.743024s => 0.003061s
=> 0.2718281828459045235360287471352662497757247093699959574966967627724076630353547594571382178525166427e1

if frac_part.zero? && y.exponent < Math.log(prec) * 5 + 20
  # Use exponentiation by squaring
else
  # Use exp(y*log(x))
end

Appropriate threshold of y.exponent depends on exp, log, and multiplication performance.
This condition is tuned with prec <= 1000000 and #407 #433 applied.
Theoretically, the condition should be y.exponent > coef * log(prec)**a where a is between 1..3, I guess.