Merge path doesn't zero out NaNs when beta=0

[brian-kelley]

In spmv (y := alphaAx + betay) if y contains a nan on input and beta is zero, all its entries should be overwritten with zero rather than multipled by zero (preserving nans) before computing alpha*Ax.

The default native algorithm seems to do this, but merge path doesn't.
We should also test this systematically in the spmv unit tests.

[brian-kelley]

I think this is actually the same issue as #2010 which was closed with a workaround, but I'll just leave this open instead of reopening that one

[cwpearson]

I think this is the offending line:

kokkos-kernels/sparse/impl/KokkosSparse_spmv_impl_merge.hpp

Line 312 in b87dc95

KokkosBlas::scal(y, beta, y);

If I'm reading scal correctly, it just just y = beta * y for this case, so if y is NaN it won't work.

The Test_Sparse_spmv_bsr.hpp file generates random vectors for tests, but those vectors do not include NaN, which is why this passes the tests.

[lucbv]

Then we probably want to have an if statement to check for beta == 0 and call Kokkos::deep_copy(y, 0); in that case