Do not cache import of numpy in C++

[jl-wynen]

The old code triggered a segfault in ScippNeXus with Python 3.10. I don't know exactly why, this is hard to reproduce. But I suspect it has to do with the tear down order. Numpy might be deallocated before scipp which creates a dangling reference with the static variable.

I benchmarked this and got

old: 0.008s
new: 0.011s

with

import scipp as sc
from timeit import repeat
ts = repeat("var.value",
            setup="var=sc.scalar(123)",
            globals={"sc": sc},
            number=10_000,
            repeat=1_000)
print(min(ts))