Relax precision for Binomial test

This should fix failure found in #166:

   log probabilty check: FAIL
     *** Failed! Falsified (after 34 tests):
     binomial 79 9.83897381493678e-2
     7
     probability    = 0.14935400356648396
     logProbability = -1.9014359280741342
     log p          = -1.9014359280741266
     eps            = 3.970429114115656e-15

It turns out that main contribution to error in probability is (1-p)^(n-k) and
for logProbability we subtract values of same magnitude with usual consequences
for precision

tests/Tests/Distribution.hs

@@ -351,7 +351,7 @@ instance Param ChiSquared where
 
 instance Param BinomialDistribution where
   prec_discreteCDF _ = 1e-13
-
+  prec_logDensity  _ = 48
 instance Param CauchyDistribution where
   -- Distribution is long-tailed enough that we may never get to zero
   cdfLimitAtNegInfinity_enabled _ = False