Exclude self-matches and correct N in corr_dim

[CSchoel]

In #20, I noted that my implementation of corr_dim deviates from the specification in two details:

• The GP article excludes self-matches, which nolds currently does not.
• The N used for normalizing the sum is the number of vectors, whose distances are compared. However, in nolds I use the length of the input array, not the number of orbit vectors. This should not introduce a large error, but should be corrected nonetheless.

Apparently, I just forgot to implement the required changes. This issue is here for reminding me to not do that again. 😉

[CSchoel]

This is fixed by #32.

It turns out the change for N didn't change the results at all, since it only introduced an offset in the y-axis in the log-log plot. Excluding the self-matches did introduce changes, but not large enough to affect any existing unit tests.

To make sure that corr_dim is correct, I added a test involving the Lorenz system. It turns out that corr_dim is quite sensitive to the lag of the embedding dimension, so I added it as another parameter.