native: bad computation of work matrix size in Dormqr

[vladimir-ch]

Dormqr computes the maximum number of columns of a work matrix nwxnb that fits into available workspace. If lwork is not long enough to hold the optimum block size nb, nb is reduced. The code however doesn't update ldwork, leaving it at the previous, unnecessarily large value. If the workspace is long enough to use the blocked code but not long enough to hold the nwxnb matrix, Dlarfb will panic later.

    ldwork := nb // nb is the optimal block size as returned from Ilaenv.
    if nb > 1 && nb < k {
        iws := nw * nb
        if lwork < iws {
            nb = lwork / nw
            // ldwork = nb // <--- Uncommenting fixes the issue.
            nbmin = max(2, impl.Ilaenv(2, "DORMQR", opts, m, n, k, -1))
        }
    }

Speaking of Dormqr, why is t allocated in Dormqr and not taken from work as in the reference LAPACK, @btracey ?

[btracey]

Um?

Here is the PR: #17

This is the first large chuck of lapack I did, and I still didn't have a great feel for how it all worked yet. It's almost certainly wrong to allocate t, and I probably did it because of the ldwork issue.