Improve IMTL-G implementation

[PierreQuinton]

Doing pseudo inverse then multiplying is equivalent (and less efficient) than solving a linear system.

In IMTL-G, we have the following code: https://github.com/TorchJD/torchjd/blob/ab9208f681c9b91a31fd667c5ca9c74db46a1d59/src/torchjd/aggregation/imtl_g.py#L45

It is equivalent to, and less efficient (and precise) than

raw_weights = torch.linalg.lstsq(matrix @ matrix.T, d)

If we do that, we do not need the try/catch anymore. Note that we need to be careful about the driver.

[ValerianRey]

I agree, based on this source: https://pytorch.org/docs/stable/generated/torch.linalg.pinv.html

However, I think you forgot the .solution in your suggested code, so it should be:

raw_weights = torch.linalg.lstsq(matrix @ matrix.T, d).solution

[ValerianRey]

Apparently, the try-catch is still necessary: matrices with values larger than like 1e18 lead to RuntimeError when applying lstsq.