That's bit of a circular argument.
Agreed that compute code for ML is not written in Python, but if GIL wasn't a problem there would be one less reason to drop down to C. Many users would appreciate that. Cython helps with the C part but to a degree.
You are forgetting python is (depending on the benchmark) still 100-400x slower than a language compiled to native machine code. Removing the GIL might slightly help in some cases but anything CPU bound requiring
high performance will still need to drop down to something faster.
The compute code for ML is emphatically not written in Python – Python is essentially a lightweight API wrapper around it.