Fix cardinality tests (#14)

* Compare approximated cardinalities to the expected errors

Since we have information about the standard error only, we
need to estimate the average relative error in order to
prove that the algorithms behave close to theoretical
expectations.

* Test HLL on larger number of unique elements

HLL was designed to behave better in the middle range of its
supported values. Since we have only one run in the test,
we need to try to use such values in order to receive
closer values to the expected ones.

* Exclude small cardinalities from the counter for error estimation

* Use relative error instead of absolute relative error