ChainerX, or chainerx is a standalone Python package that's integrated into Chainer.
It is implemented almost purely in C++ where the C++ implementations are exposed via Python bindings, allowing Chainer to make use of it.
It does in other words not replace Chainer. It aims to instead improve the performance of Chainer in terms of speed by reducing the Python overhead.
Following guides ares aimed toward users familiar with the Chainer interface but want to improve their training/inference speed, using ChainerX. They explain how to install it, the motivation behind it and how to modify your existing Chainer code to use ChainerX. The last section is the product backlog of ChainerX listing its milestones.
See ChainerX tutorial.
See: ChainerX limitations
Yes, it is possible. See the code samples below.
It is almost identical to the Python interface with a 1-to-1 mapping. The binding layer is thin and many of the types defined in C++ have Python equivalent classes. The bindings are defined here.