Unify coerce prec calculation

[tompng]

Fixes #388
Unify coerce logic in add/sub/mult/div/coerce/cmp.
There are some differences. Unify them to div's precision calculation because add/sub/mult's calculation coerce Rational with low precision.

# div
BigDecimal(1) / (3/7r) # Same as `BigDecimal(1) / BigDecimal(3/7r, 0)`
# => 0.233333333333333333333333333333332555555555555556e1 # ok
BigDecimal(1).div(3/7r, 50) # Same as `BigDecimal(1).div(BigDecimal(3/7r, 50), 50)`
# => 0.23333333333333333333333333333333333333333333333333e1 # ok

# add, sub, mult
BigDecimal(1) + (3/7r)
# before: 0.1428571429e1 (too low)
# after: 0.142857142857142857142857142857143e1 
BigDecimal(1).add(3/7r, 50)
# before: 0.1428571429e1 (too low)
# after: 0.14285714285714285714285714285714285714285714285714e1

Cmp with rational will change. The original behavior was weird.

10/3r == BigDecimal(10/3r, prec)
# before: always false
# after: always false

1/3r == BigDecimal(1/3r, prec)
# before: true when prec == 9*n (n>=1)

# after: true when prec == 9*n (prec >= 2*double_fig)