added the computed initialization of the jac approx in Broyden

[zaccharieramzi]

WIP: need to run the existing tests and add the previously failing test written by @vroulet in a previous PR

In order to implement this change, I had to change the dtype of gamma which is now a pytree rather than a scalar, since it can be different for each leaf of the parameters.

[zaccharieramzi]

So actually the failure that @vroulet was showcasing here is actually already a test in the broyden test suite.

Everything is green now (with simpler settings), so this is ready for review.