Reformulates the section on cross-platform portability of pickles,

• pickles are generally portable across platforms, and in rare cases when estimators aren't (because of custom low level structures) I think we should consider it as a bug and fix when possible (as was done in Fixed a cross-platform endian issue #17644 for instance). It shouldn't be too difficult, particularly that we don't have backward compatibility constraints on pickles. Though the issue is more that it's difficult to test this as part of CI. Otherwise there is no way to deploy a model trained on x86_64 to arm64, aside maybe from using ONNX which has other constraints.
• Docker cannot be used for cross platform portability but is rather used to freeze the environment for a given architecture.

cc @ogrisel