HM is not complex type inference. In fact, among all the approaches you cite, it leads to the simplest type system and the simplest implementation. Moreover, there are lot's of courses, reference implementations, and reasonable extensions for a wide array of features (structural subtyping, qualified types, etc). There are even type-system libraries to make it easy to implement (like Inferno).

When new programmer discover ML-family language, they are often stunned by how lightweight it feels, because you don't have to annotate anything. If your types are not structurally too complicated and you want something really easy to use, HM is still the nicest experience.

Naturally, it's all a tradeoff, and if you want specific features (chiefly: borrows, high-order types, overloading, or Typescript-like features), you will need to abandon complete inference (and use something like bidirectional, most likely).