Fix swapped location indices in UnivariateMaternNuggetsStationary covariance

[fonzie42]

Summary

UnivariateMaternNuggetsStationary::GenerateCovarianceMatrix passes the two
location indices to CalculateDistance in the wrong order, (j0, i0) instead
of (i0, j0). Every other kernel in the project passes (i0, j0).

Root cause

The distance helper signature is:

CalculateDistance(Locations &loc1, Locations &loc2, idxInto_loc1, idxInto_loc2, ...)

In the generation loop, i0 walks the rows (indexing aLocation1) and j0
walks the columns (indexing aLocation2), so the correct call is
(..., i0, j0, ...). The nuggets kernel instead calls (..., j0, i0, ...) in
both branches.

This is invisible for the symmetric train-train covariance matrix, because
the Euclidean distance is symmetric: d(loc[i], loc[j]) == d(loc[j], loc[i]).
That is why synthetic data generation and MLE are unaffected, and why the
existing kernel test (which only checks the symmetric matrix) never caught it.

It does corrupt the rectangular cross-covariance matrix C12 assembled
during prediction, where the two location sets (observed and missing) differ.
Entry (i, j) ends up using d(location1[j], location2[i]) instead of
d(location1[i], location2[j]), the transposed matrix, producing wrong
kriging predictions whenever this kernel is used for prediction.

Fix

-                expr = DistanceCalculationHelpers<T>::CalculateDistance(aLocation1, aLocation2, j0, i0, aDistanceMetric,
+                expr = DistanceCalculationHelpers<T>::CalculateDistance(aLocation1, aLocation2, i0, j0, aDistanceMetric,

(applied to both branches of GenerateCovarianceMatrix).

Test

Adds TestUnivariateMaternNuggetsCrossCovariance.cpp: it builds a 3×3
cross-covariance between two distinct location sets and checks every entry
against an independently computed reference. With nu = 0.5 the Matérn
covariance reduces to the exponential covariance sigma^2 * exp(-d / beta), so
the reference values are self-evident and do not depend on any internal index
convention.

• With the swapped indices, entry (0,1) resolves to exp(-1) ≈ 0.36788 (the
  transposed pair, distance 1) instead of exp(-3) ≈ 0.04979 (distance 3), so
  the test distinguishes the two index orders.
• With the corrected indices, every entry matches the reference.
• The existing symmetric kernel test is unaffected, since the symmetric
  (data-generation / MLE) path does not depend on the index order.

Note

Companion fix: the R-interface prediction problem-size fix (#59). The
two are independent code changes (disjoint files); both are needed together for
fully correct end-to-end kriging predictions with the nuggets kernel.