Multiple taichi instances(backends) in single process.

[iperov]

Current version of Taichi can run only single backend in single process.
This imposes serious restrictions on the parallelization of calculations using multiple CPU/GPUs of different backends.

In this case, it is necessary to run separate python processes.
However, there are disadvantages to this:

1. each python process eats up an additional couple of GB of swap
2. The data must be transferred between processes, which slows down the computation

You may need to change the architecture of the Taichi.
But it's better to start now, and don't waste time.