- The use of 'functorch.compile' over a function that takes variable-size
  batch inputs proves impossible, as the tracing on first call creates a
  computation graph with fixed dimensions.
- As a result, the tentative compilation of per-sample clipped gradients
  computation prevents the proper use of DP-SGD with the functorch backend.
- An alternative attempt was to compile the sample-wise function and vmap
  it afterwards, but this is currently unsupported (and unlikely to be as
  part of functorch, as development efforts have moved to 'torch.func').
- This commit therefore drops the use of 'functorch.compile'.

Improve tests coverage and fix test-digged bugs

See merge request declearn2!57

Until now, when provided with y_true:<shape>, y_pred:(<shape>, 1)
inputs, the MAE, MSE and R2 metrics would be entirely wrong, due
to the way numpy casts operations between such inputs. This patch
adds some shape-verification and squeezing operations that fix the
computations. Unit tests were added to cover this case.

Implement 'L2GlobalClipping' OptiModule.

See merge request !56