No, you can be an ML practitioner with just an intuitive understanding of, say, gradient descent works and you would do fine. You can even pick up that intuitive understanding on a strictly need-to-know basis, when it's needed for learning an ML technique. That's what fast.ai teaches.
For being more than a practitioner, like an implementer of new ML libraries or a researcher, of course you'd need to know more.
No, but there are some good fundamentals. I.e, The optimization bit for when dealing with SVM, Kernel methods, etc. All parts that college ML courses cover in depth
