Fix BigMath.atan precision safe margin

[tompng]

Increase BigMath.atan(x, prec) precision when |x|>0.5

BigMath.atan(x, prec) is calculated with precision=prec+BigDecimal.double_fig when |x| <= 0.5.
But this safe margin was not properly applied when |x| > 0.5.

BigMath.atan(BigDecimal('0.2'), 100) #=> accuracy was about 116 digits
BigMath.atan(BigDecimal('0.8'), 100) #=> accuracy was about 108 digits
BigMath.atan(BigDecimal('3'), 100) #=> accuracy was about 100 digits

Added test code that failed before

atan(BigDecimal("0.8"), n) # Precision is not enough: 107 < 100 + 12.
atan(BigDecimal("3"), n) # Precision is not enough: 50 < 50 + 12.