Looks good, but more work is needed.

The case of a suboptimal memory allocator - is the problem of shared linking. Ideally, we should make a "self-contained" shared library. It means that every dependent library, including malloc, is statically linked into it, and the set of exported symbols is limited to a minimum. But I'm unsure if it will work with Python - never used pybind.

This comment is important :)

Noted, thanks.

Progress is happening! ~24.8% performance increase and much better results expected on the weak tests with the latest jemalloc patches by @auxten + a golang binding as alternative to the python one for those afraid of snakes! ;)