Unexpect output with dplasma_dlange for all-zero matrix

[Qiao-Paper]

Describe the bug

For all-zero matrix, its dplasmaFrobeniusNorm calculated by dplasma_dlange is -nan.

To Reproduce

Steps to reproduce the behavior:

1. Initialize a all-zero matrix
2. Calculate its FrobeniusNorm: dplasma_dlange( parsec, dplasmaFrobeniusNorm, A);

Expected behavior

The result should be 0 instead of -nan.

Bug Location

When a new tile is considered during the Frobenius norm calculation, its norm is compared with scale and divided (the smaller value divided by the larger one). In this way, 0/0 division occurs when all the tile is 0.

[QingleiCao]

I do not like the current solution for Norm. We can use an additional buffer and remove the complicated flows.

[therault]

I discussed with @mfaverge about this:

• the algorithm may seem more complicated than needed, but the scalings are here to ensure numerical stability
• there is a check missing in a couple of places. When scaling, one should scale only if the absolute value of the divisor is not 0, as your test show. It is done in the kernels plssq and gessq, but when we do that inline we should also check (see https://gitlab.inria.fr/solverstack/chameleon/-/blob/master/coreblas/include/coreblas/sumsq_update.h?ref_type=heads#L50 for example)
• All the NEW are to allocate local norms that are reduced, everything is done in DAG / PTG, and yes, if performance are a must on this, combining PTG and MPI to do local computations on a workspace and then reducing the result with MPI would probably be more efficient.

@mfaverge said he will provide a PR to fix the missing checks.