nbdev_test parallel() process re-use results in hard to debug behaviour

Run `nbdev_test --n_workers 1 --file_re "0[07].*" --do_print` in `fastai` repo (after installing the dev dependencies)
Error I'm getting:
```
xl0@vespa:~/work/code/fastai$ nbdev_test --n_workers 1 --file_re "0[70].*" --do_print
Starting /ssd/xl0/work/code/fastai/nbs/00_torch_core.ipynb
- Completed /ssd/xl0/work/code/fastai/nbs/00_torch_core.ipynb
Starting /ssd/xl0/work/code/fastai/nbs/07_vision.core.ipynb
AttributeError in /ssd/xl0/work/code/fastai/nbs/07_vision.core.ipynb:
===========================================================================

While Executing Cell #41:
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[1], line 1
----> 1 test_fig_exists(ax)

File ~/mambaforge/envs/test/lib/python3.9/site-packages/fastcore/test.py:101, in test_fig_exists(ax)
99 def test_fig_exists(ax):
100 "Test there is a figure displayed in `ax`"
--> 101 assert ax and len(ax.figure.canvas.tostring_argb())

File ~/mambaforge/envs/test/lib/python3.9/site-packages/matplotlib/backends/backend_agg.py:438, in FigureCanvasAgg.tostring_argb(self)
431 def tostring_argb(self):
432 """
433 Get the image as ARGB `bytes`.
434
435 `draw` must be called at least once before this function will work and
436 to update the renderer for any subsequent changes to the Figure.
437 """
--> 438 return self.renderer.tostring_argb()

AttributeError: 'FigureCanvasAgg' object has no attribute 'renderer'

- Completed /ssd/xl0/work/code/fastai/nbs/07_vision.core.ipynb

nbdev Tests Failed On The Following Notebooks:
==================================================
07_vision.core.ipynb
(test) xl0@vespa:~/work/code/fastai$
```

` --file_re "0[07].*"` meaning, test notebooks 00 and 07.
nbdev_test will use `fastcore` `ProcessPoolExecutor` with 1 worker, which means the worker process will be re-used between notebooks.

I don't completely understand the source of the error - both notebooks 00 and 07 make use of `test_fig_exists(ax)`, but only one of them fails, but pretty certain it comes from some internal state being shared between notebooks by the re-used process: No failure occurs if notebooks are run separately, or if `--n_workers 2` and each notebook gets a fresh worker process.

While the issue could possibly be fixed in `fastai`, I believe that the behaviour is itself a bug - a test framework should not leak state between unrelated notebooks. This behaviour results in unexpected and frustrating failures. I earlier ran into a similar issue using JAX with nbdev, when two notebooks would initialize JAX with different parameters, but since the process is being re-used, the second initialization in the same process is ignored.