Don't assume strings are valid unicode

The logic in conversions.jl:61 seems to assume that the Julia string being converted to a Python string is valid unicode. Some Python libraries (eg tensorflow) use the Python string type for a binary blob that might not be valid unicode.

Start at conversions.jl:61 and inspect how Julia strings are converted to Python strings. Reproduce the issue with a binary value that is not valid Unicode, such as the kind of blob used by TensorFlow libraries. Done means conversion no longer assumes every string is valid Unicode and preserves binary data correctly.

Rédigé par le modèle d'indexation à partir du texte de l'issue.