Add a workaround for slow BigDecimal#to_f when it has large N_significant_digits

[tompng]

BigDecimal#to_f is accurate, it returns exact nearest float value, but may be extremely slow in some case.
BigMath methods needs workaround like this:

irb(main):003> x = BigDecimal(1).div(7, 10000); 10000.times{(x - it).to_f}
processing time: 1.746866s
=> 10000
irb(main):004> x = BigDecimal(1).div(7, 10000); 10000.times{(x - it).mult(1, 20).to_f}
processing time: 0.053746s

In my environment, rake test time reduced from 4.8..4.9 seconds to 4.5..4.7 seconds.

The root cause is not fixed. It'd be better to optimize BigDecimal#to_f in the future.