perf: More tse optimisations

Benford Correlation is once again slightly faster now because I replaced value_counts by unique_counts. By putting pl.int_range(1, 10) in front, we do not need to sort, and unique_counts counts in the given order. In addition, unique counts is just slightly faster than value_counts because it only counts and does not return the corresponding values.
Number Crossing. Simplified the computation and got a significant speed boost.
percent_reoccurring_points,. Simplified the mathematical formula.
sum_reoccurring_values. The original one-line implementation was elegant, but is_unique, unique, and filter are more expensive. We can directly do a group_by (value_counts), which cuts down the size of the series, and then filter.. This is 40-50% faster both when there are lots of unique values and when there are almost no unique values.

Clean Up and Docs
1. Some features will return a column called "literal" in lazy mode. I fixed that in tests. In the name space, we should add suffix.
2. Cleaned up docstrings here and there.

functime-org / functime