Some easy micro optimizations to collection

[bluetech]

Collectively, on this benchmark run with --collect-only:

import pytest
@pytest.mark.parametrize("x", range(50000))
def test_foo(x): pass

Before:

============================ no tests ran in 11.37s ============================
         23775445 function calls (22920574 primitive calls) in 11.648 seconds

After:

============================ no tests ran in 10.51s ============================
         21725509 function calls (20870638 primitive calls) in 10.931 seconds

The first commit utilizes the fact that dicts are insertion-order preserving on Python>=3.7 to avoid OrderedDict overhead when possible.

The second and third commits apply some micro optimizations to Node.

The fourth applies some minor optimizations to PythonCollector.collect.

[bluetech]

I scaled back the last commit a bit after the CI kindly informed me that I had missed some details 😄 Doesn't affect the benchmark results however.

[bluetech]

cc @boxed, #5516

[boxed]

I had some time to check this and the other changes that have landed in master. I am comparing 5.4.1 and master:

It's slightly better but it's still very bad.

[boxed]

I tried pytest 3.10.1 and it's at 7.8s for the parametrize_huge benchmark. Again still very bad, but a lot better than pytest 5. I even checked 3.6 because I have a memory of this being faster on our suite at work: 7.5s so basically the same.

Nose2 beating out hammett on parametrize_huge is pretty impressive. Don't really understand how :P