Permit inf samples in ParameterBijectorsTest.

During Hypothesis testing in ParameterBijectorsTest, some distributions (such as `GammaGamma`) can generate samples too large to be represented in float32.  This change updates the test to just check that samples are non-non, instead of checking that all samples are finite.

PiperOrigin-RevId: 341067297

tensorflow_probability/python/distributions/distribution_properties_test.py

@@ -341,8 +341,8 @@ def testCanConstructAndSampleDistribution(self, data):
     x = self.evaluate(new_dist.sample(seed=test_util.test_seed()))
     self.assertEqual(sample_shape, x.shape)
 
-    # Valid parameters should give finite samples.
-    self.assertAllFinite(x)
+    # Valid parameters should give non-nan samples.
+    self.assertAllFalse(np.isnan(x))
 
 
 def _all_shapes(thing):