Error about mixing pipelines in nosetests

When testing a PTranform (defined using @ptransform_fn) that merges several PCollections from different sources the following error is raised:
```
ValueError: Mixing value from different pipelines not allowed.
```

Actually running the same pipeline in GCP using the `DataflowRunner` does not give any error. Neither does running the test file manually instead of through nose.

Here is an example:
```

# Defined in module `utils`
@ptransform_fn
def Join(pcolls, by):
return pcolls | beam.CoGroupByKey()

```

```

class UtilsTest(unittest.TestCase):
def test_join(self):
p = TestPipeline(runner="DirectRunner")

p1 = (p
| "Create p1" >> beam.Create([
{'a': 1, 'b': 11},

{'a': 2, 'b': 22},
{'a': 3, 'b': 33}]))
p2 = (p
|
"Create p2" >> beam.Create([
{'a': 1, 'c': 111},
{'a': 1, 'c': 112},

{'a': 3, 'c': 333}]))

res = ((p1, p2) | "LeftJoin" >> utils.Join(by='a'))

beam.assert_that(res, beam.equal_to([
{'a': 1, 'b': 11, 'c': 111},
{'a':
1, 'b': 11, 'c': 112},
{'a': 2, 'b': 22},
{'a': 3, 'b': 33, 'c': 333}]))

# Run test pipeline
p.run()

```

Imported from Jira [BEAM-1996](https://issues.apache.org/jira/browse/BEAM-1996). Original Jira may contain additional context.
Reported by: while.

Start with the supplied UtilsTest.test_join reproduction, running it through nose and manually with DirectRunner to compare the behavior. Trace utils.Join, the @ptransform_fn wrapper, and beam.CoGroupByKey to identify why PCollections from the same pipeline appear mixed; done means the nose run no longer raises the cross-pipeline ValueError while the existing join assertions still pass.

Written by the indexing model from the issue text.