DESCRIPTION

Add a PEP 561 type stub (special.pyi) for numba.misc.special to fix a false-positive not-iterable diagnostic emitted by static type checkers (ty 0.0.49+, pyright 1.1.410+, mypy) on prange loops.

PROBLEM

prange is implemented as a class whose __new__ delegates to range:

# numba/misc/special.py
class prange(object):
    def __new__(cls, *args):
        return range(*args)

At runtime, prange(n) returns a standard range object which is fully iterable, but the mianstream type checkers don't know it, as they see prange as a class and infer that calling it returns a prange instance, which has no __iter__ method. This triggers a false-positive error on every for i in prange(n): loop:

error[not-iterable]: Object of type `prange` is not iterable
  --> evaluate.py:1818
   |
...  for i in prange(n_sites):
   |          ^^^^^^^^^^^^^^^
   |          It doesn't have an `__iter__` method or a `__getitem__` method

Users then have to add # type: ignore[not-iterable] on every prange line, which is noisy and masks real bugs.

FIX

This PR adds numba/misc/special.pyi with three @overload signatures for prange, matching range's API:

@overload
def prange(stop: int, /) -> range: ...
@overload
def prange(start: int, stop: int, /) -> range: ...
@overload
def prange(start: int, stop: int, step: int, /) -> range: ...

pndindex is also stubbed for completeness (-> np.ndindex).

The existing .pyi files in numba (e.g. _helperlib.pyi, core/decorators.pyi) follow the same PEP 561 pattern, so this is consistent with numba's existing approach to type stubs.

TEST

None unit test is prepared since it's static and won't hurt any other functions. It has been manually verified with ty 0.0.49 and pyright 1.1.410 that the not-iterable diagnostic is resolved on the newest numba release.