More torch fixes

[honno]

• Remove use of _UndefinedStub in dtype_helpers.py
  • This is mighty useful as currently we have two classes of the dtype helpers—those that consume stubs and those who don't, which is hard to keep track. This unifies the approach of the helpers so they are only constructed with available dtypes.
• Resolves Wrong scalar type used in complex tests #179
• Resolves test_int_to_dtype meta-test fails with torch #178 by introducing a hh.reject_overflow() context manager that deals with non-OverflowError overflow errors (i.e. torch's RuntimeError: ...overflow...)
• Resolves Tests for comparison functions should not use complex dtypes #180
• Resolves conj, real, and imag not found in array_api_compat.torch #186
• Resolves iinfo() tests should ignore undefined unsigned dtypes #166

[asmeurer]

Here's another error:

>       return torch.full(shape, fill_value, dtype=dtype, device=device, **kwargs)
E       RuntimeError: value cannot be converted to type c10::complex<float> without overflow
E       Falsifying example: test_full(
E           shape=(),
E           fill_value=complex(0.0, 3.402823466385289e+38),
E           kw={},
E       )

full uses the default complex dtype by default, which is complex64 for torch, so it shouldn't generate fill values outside of that dtype unless dtype=complex128 explicitly.

[asmeurer]

There's also this error. I can't tell if it's caused by this or something else:

_______________________________________________________________________________ test_prod _______________________________________________________________________________

    @given(
>       x=xps.arrays(
            dtype=xps.numeric_dtypes(),
            shape=hh.shapes(min_side=1),
            elements={"allow_nan": False},
        ),
        data=st.data(),
    )

array_api_tests/test_statistical_functions.py:112:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
array_api_tests/test_statistical_functions.py:143: in test_prod
    d_m, d_M = dh.dtype_ranges[default_dtype]
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = EqualityMapping({torch.int8: MinMax(min=-128, max=127), torch.int16: MinMax(min=-32768, max=32767), torch.int32: MinMa...86e+38, max=3.4028234663852886e+38), torch.float64: MinMax(min=-1.7976931348623157e+308, max=1.7976931348623157e+308)})
key = None

    def __getitem__(self, key):
        for k, v in self._key_value_pairs:
            if key == k:
                return v
        else:
>           raise KeyError(f"{key!r} not found")
E           KeyError: 'None not found'
E           Falsifying example: test_prod(
E               x=tensor(0, dtype=torch.uint8),
E               data=data(...),
E           )
E           Draw 1 (kw): {}

array_api_tests/dtype_helpers.py:83: KeyError

[asmeurer]

I think that's it. Every other test failure seems to be either an issue with torch or an unrelated test suite problem which I've reported elsewhere.

[asmeurer]

Not that it's that important, but can we address #181 (comment) here?

[honno]

full uses the default complex dtype by default, which is complex64 for torch, so it shouldn't generate fill values outside of that dtype unless dtype=complex128 explicitly.

There's also this error. I can't tell if it's caused by this or something else:

Fixed and fixed—thanks for catching these. I folded in #182 as it addresses the former issue. ^{Whilst its nice to keep PRs atomic I think we're not keeping track of multiple PRs at once, so mono PRs with at least a general cohesion is probably better.}

Not that it's that important, but can we address #181 (comment) here?

Already addressed!

[honno]

(I'll merge this soon if no review)