Is it merely a coincidence that certain modules' numba operations respect PISA_NUM_THREADS?

[thehrh]

Several services implement numpy generalised universal functions using from numba import guvectorize.
Unless I'm misunderstanding, numba.set_num_threads(PISA_NUM_THREADS) in https://github.com/icecube/pisa/blob/master/pisa/utils/numba_tools.py#L71 is always executed (for the parallel target) only after some involved chain of imports.

For example, take stages/xsec/dis_sys.py, which does not explicitly import from pisa.utils.vectorizer unlike other services (e.g., stages/pid/shift_scale_pid.py).
I believe we reach the line above via from pisa.core.stage import Stage -> from pisa.core.container import ContainerSet -> from pisa.core.translation import histogram, lookup, resample -> from pisa.utils import vectorizer -> from pisa.utils.numba_tools import cuda_copy 😅

Since all services import the Stage base class, it always seems to work out fine, but shouldn't we attempt to call numba.set_num_threads already upon initialising the pisa package (in particular, given that numba is an install requirement now)?