RFSurrogates

This PR fixes an issue where the result of `var.relations` would contain
long runs of NA starting in the latter half of variables. This issue
caused mean adjusted agreements to be shifted horizontally and across
following rows.

The issue was caused by setting the mean adjusted agreement of a
variable with itself to NA incorrectly. Instead of using the current
index, the variable index was used. This resulted in cells being
incorrectly overwritten as NA if the variable index was smaller than the
number of all variables, and in the implicit addition of NA beyond the
end of the vector until the variable index when the variable index was
larger than the number of all variables.

This caused `mean.index` to return vectors with a length not exactly
equal to the number of variables, which in the following steps caused an
overrun when the concatenated vectors were being cast to a matrix with
dimensions being exactly the length of all variables by all candidate
variables. A warning message would appear when this occurred: `data
length [...] is not a sub-multiple or multiple of the number of rows
[...]`. It did not however notify that extra data was dropped
implicitly.

This PR also contains some basic test cases for the internal
`mean.index` to prevent regression of this issue.
- Assert that the length of the return value is always exactly as long
as the number of variables
- Assert that for unknown variables (that is, variables not represented
in `list.res` or `index.variables`) the function will return an NA
vector of the appropriate length.