In #713 and #1296, the `google` package in protobuf sources was found
to cause conflicts with other Google projects, because it was not
properly configured as a namespace package [1]. The initial fix in
786f80f addressed part of the issue, and #1298 fixed the rest.
However, 786f80f (the initial fix) also made `google.protobuf` and
`google.protobuf.pyext` into namespace packages. This was not correct:
they are both regular, non-namespace, sub-subpackages.
However (still), the follow-up #1298 did not nominate them as
namespace packages, so the namespace registration behavior has
remained, but without benefit.
This change removes the unnecessary namespace registration, which has
substantial overhead, thus reducing startup time substantially when
using protobufs.
Because this change affects the import internals, quantifying the
overhead requires a full tear-down/start-up of the Python interpreter.
So, to capture the full cost for every run, I measured the time to
launching a _fresh_ Python instance in a subprocess, varying the
imports and code under test. In other words, I used `timeit` to
measure the time to launch a _fresh_ Python subprocess which actually
performs the imports.
* Reference: normal Python startup (i.e., don't import protobuf at all).
```
% python3 -m timeit -s 'import subprocess' -r 3 -n 10 'subprocess.call(["python3", "-c", "pass"])'
10 loops, best of 3: 27.1 msec per loop
```
* Baseline: cost to import `google.protobuf.descriptor`, with
extraneous namespace packages.
```
% python3 -m timeit -s 'import subprocess' -r 3 -n 10 'subprocess.call(["python3", "-c", "import google.protobuf.descriptor"])'
10 loops, best of 3: 133 msec per loop
```
* This change: cost to import `google.protobuf.descriptor`, without
extraneous namespace packages.
```
% python3 -m timeit -s 'import subprocess' -r 3 -n 10 'subprocess.call(["python3", "-c", "import google.protobuf.descriptor"])'
10 loops, best of 3: 43.1 msec per loop
```
[1]: https://packaging.python.org/guides/packaging-namespace-packages/
