Red2Band: revisit distributed panel computation to group communications

[albestro]

This PR explores a possible optimization of reduction to band panel computation, by grouping the two all reduce in a single one (with more data) in each reflector computation.

Panel computation without this PR

for (j : panel.cols()) {
  computeX0AndSquares();    // each rank computes local sum(squares) + just head rank extract x0

  allReduce(x0, squares);   // each rank get total sum(squares) + x0 (it just get broadcasted)

  tau = computeReflector(); // now the reflector can be computed
  computeW();               // each rank computes w = Pt* . v

  allReduce(w);             // each rank get the final w

  updateTrailingPanel();    // update trailing panel with v and w
}

With this PR instead of having two separate MPI_AllReduce we group them in a single one (with more data), and it needs a bit of reshuffling of the operations

for (j : panel.cols()) {
  computeX0AndSquares();             // each rank computes local sum(squares) + just head rank extract
  computeWtmp();                     // each rank start computing w, but first row is skipped
  
  allReduce(x0, squares, row0, w_s); // x0, sum(squares), row0, sum(w0)

  tau = computeReflector();
  computeW();                        // w = row0.T + (1 / (x0 - y)) * w
  updateTrailingPanel();             // update trailing panel with v and w
}

• Red2Band: use well-formed reflector for internal panel computations #1348 this work requires the refactoring happened there
• Evaluate use of std::span (see Red2Band: use well-formed reflector for internal panel computations #1348 (comment))

[albestro]

cscs-ci run