hypothesis.extra.numpy.from_dtype is difficult to use with complex dtypes

[felixdivo]

The following code ...

import numpy
from hypothesis.extra.numpy import from_dtype

print(repr(from_dtype(numpy.dtype("complex64"), max_magnitude=3).example()))

... generates this exception with varying values (line breaks added):

hypothesis.errors.InvalidArgument:
min_value=-2.958039891549808 cannot be exactly represented as a float of width 32
- use min_value=-2.9580399990081787 instead.

It is raised here (with a useless strack trace):

hypothesis/hypothesis-python/src/hypothesis/strategies/_internal/numbers.py

Lines 488 to 492 in 68b9c20

	if min_value != min_arg:
	raise InvalidArgument(
	f"min_value={min_arg!r} cannot be exactly represented as a float "
	f"of width {width} - use min_value={min_value!r} instead."
	)

Curiously, the same setup works for hypothesis.strategies.complex_numbers:

from hypothesis.strategies import complex_numbers

print(repr(complex_numbers(max_magnitude=3).example()))

Tested on hypothesis 6.67.1.

[felixdivo]

Thanks @honno for looking into this!