Avoid to compute the pseudoinverse to evaluate the external wrenches

[GiulioRomualdi]

According to https://eigen.tuxfamily.org/dox/group__LeastSquares.html this should be more efficient

cc @traversaro @S-Dafarra

[traversaro]

I just realized that in most cases we can also avoid the Qr decomposition, as in the submodels has only one full wrench as unknown, in the equation Ax = b the A is just an adjoint matrix for which the inverse is trivial. However, computing this "special case" is probably more complex that this change, so let's see how time the use of Qr saves.

[GiulioRomualdi]

I would like to present you some results. I used perf to profile the new implementation with respect to the old one (the one with pinv)
The command I used is sudo perf record --call-graph dwarf -p <PID> -- sleep 3 then I analyzed the data with hotspot

The following figures shows the result in case of pinv (the old method) and the qr decomposition (what the PR is proposing)

pinv	QR decomposition

Accordingly to the results the approach I proposed is 2 times faster than the original one.

[traversaro]

Great, let's merge then. Given that the base branch of the PR for some reason is quite old, can I rebase on top of the latest master to run the CI?

[GiulioRomualdi]

Great, let's merge then. Given that the base branch of the PR for some reason is quite old, can I rebase on top of the latest master to run the CI?

sure!

[traversaro]

CI is failing as we modified a structures without re-generating MATLAB bindings, I just launched the action to regenerated the bindings: https://github.com/robotology/idyntree/actions/runs/2949426070 .

[GiulioRomualdi]

cc @gabrielenava

[DanielePucci]

@gabrielenava we might consider to re-install the iDyntree for the iRonCub