Implement BigMath.erf(x, prec) and BigMath.erfc(x, prec)

[tompng]

Followup of #336

erf(x, prec)

Uses taylor series of exp(x**2)*erf(x).
Precision management is easier than normal taylor series of erf(x).

# Sign of the coefficients switches. Considering loss of significance, extra high precision is required.
erf(x) = (2/√π) * (x - x**3/3 + x**5/10 - x**7/42 + x**9/216 - ...)

# Coefficients are all positive. Need to do just what BigMath.exp(x) does.
erf(x) = (2/√π) * exp(-x**2) * (x + 2*x**3/3 + 4*x**5/15 + 8*x**7/105 + 16*x**9/945 + ...)

erfc(x, prec)

BigMath.erfc(BigDecimal(100), 10) #=> 0.64059614249217320390213389e-4345

Calculating this with 1 - erf(x, prec + extra_digits) need extra 4345 precision of erf(x). The convergence is very slow.
To calculate faster, first try asymptotic expansion and fallback to 1 - erf(x, prec + extra_digits)