Clear all n_rowsA entries in gmt_matrix_vector_mult

[Esteban82]

gmt_matrix_vector_mult cleared n_colsA entries of the output vector instead of n_rowsA.

In builds without LAPACK this left the last two values of the smoothing spline (sample1d -Fs) wrong, which is why GMT_splines, smooth and smooth_deriv fail only in the Windows CI.

Written with Claude Sonnet 5.5, reviewed with Claude Opus 5.5.

[joa-quim]

Again, not sure it's acceptable. I never had that problem in my local build but it's also true that I am using LAPACK

[seisman]

I think this is correct.

The gmt_matrix_vector_mult function is for matrix-vector multiplication $A\mathbf{b} = \mathbf{c}$.

An $m \times n$ matrix $A$ multiplies an $n \times 1$ column vector $\mathbf{b}$ should produce an $m \times 1$ result vector $\mathbf{c}$. I.e., the resulting vector $c$ should have $m$ elements (the n_rowsA variable in the C code).

It can also be verified by codes near line 1347:

			c[row] += A[ij] * b[col];

in which row is 0 - n_rowsA.

[Esteban82]

Again, not sure it's acceptable. I never had that problem in my local build but it's also true that I am using LAPACK

Indeed, with LAPACK the bug is hidden: gmt_matrix_matrix_mult clears the whole output before calling this function, so the memset inside never matters. Without LAPACK nobody does, and the last two values of the smoothing spline stay dirty.

As a check, the Windows CI (which builds without LAPACK) now passes the 3 tests that were failing because of this: GMT_splines, smooth and smooth_deriv (https://github.com/GenericMappingTools/gmt/actions/runs/37987086526).