dask.delayed memory management when a single task can consume a lot of memory outside of python

Question

I have some calculations calling the pardiso() solver from python. The solver allocates its own memory in a way that is opaque to python, but the pointers used to access th