Split floatX tests into their own CI job

[michaelosthege]

Currently our CI has a floatX job matrix on Ubuntu and Windows, running each set of tests with both float32 and float64.

But there are just a handful of tests where floatX actually makes a difference.
Only these tests should run with a floatX job matrix, thereby dramatically reducing total CI runtime.

The following screenshot is an incomplete list of candidates:

[ricardoV94]

I would definitely remove test_distributions from float32 tests. Those xfails are not informative of anything because they arise from comparing with scipy methods that are always computed on float64

[ricardoV94]

Some tests I think make sense parametrizing with both floatX:

pymc/pymc/tests/test_mixture.py

Line 662 in b895e40

def test_mixture_dtype(self):

pymc/pymc/tests/test_distributions_random.py

Lines 1577 to 1581 in b895e40

	def check_dtype(self):
	assert pm.Constant.dist(2**4).dtype == "int8"
	assert pm.Constant.dist(2**16).dtype == "int32"
	assert pm.Constant.dist(2**32).dtype == "int64"
	assert pm.Constant.dist(2.0).dtype == aesara.config.floatX

pymc/pymc/tests/test_smc.py

Line 317 in b895e40

def test_custom_dist_sum_stat(self):

pymc/pymc/tests/test_smc.py

Line 336 in b895e40

def test_custom_dist_sum_stat_scalar(self):

Probably pm.floatX helper is also tested directly, hopefully with the flags as a parametrization or something.

There should be also a DensityDist test that checks the dtype follows floatX

This list is pretty ad hoc though

[michaelosthege]

Running this on Windows indicates a previously unknown... behavior.

    @pytest.mark.parametrize("floatX", ["float32", "float64"])
    def test_dtype(self, floatX):
        with aesara.change_flags(floatX=floatX):
            assert pm.Constant.dist(2**4).dtype == "int8"
            assert pm.Constant.dist(2**16).dtype == "int32"
>           assert pm.Constant.dist(2**32).dtype == "int64"
E           AssertionError: assert 'uint64' == 'int64'
E             - int64
E             + uint64
E             ? 

Also for -(2**32) is becomes uint64.

What should I do about it?
Looks like an Aesara bug. I opened aesara-devs/aesara#871 and will reference in a conditional XFAIL.