bad autoscale_view() with long curved FancyArrowPatch

Nobody has claimed this yet.

### Bug report

When using a very long `FancyArrowPatch` with `connectionstyle="arc3"`, subsequent axes autoscaling can badly squash the figure's content. Example below exhibits the problem, but more extreme squashing can easily be obtained. Possibly this is not limited to `FancyArrowPatch`?

**Code for reproduction**

```python
import matplotlib
import matplotlib.pyplot as plt

fap = matplotlib.patches.FancyArrowPatch(
(10, 10),
(20, 2e4),
mutation_scale=50,
connectionstyle="arc3,rad=0.3",
)

_, ax = plt.subplots()
ax.add_patch(fap)
ax.set_title(f"matplotlib {matplotlib.__version__}")
ax.autoscale_view()
ax.figure.savefig("curved-arrow.png")
```

**Actual outcome**


**Expected outcome**

The axes view should span the drawn content. Any control point(s) required for the arc should be ignored.

**Matplotlib version**

* Operating system: Linux
* Matplotlib version: 3.3.4
* Matplotlib backend: Reproducible with at least `GTK3Agg` and `agg`.
* Python version: 3.9

Matplotlib installed via pip.

Start with the reproduction code using FancyArrowPatch, connectionstyle="arc3", and ax.autoscale_view(), then trace the FancyArrowPatch and autoscale_view entry points. Confirm the behavior with the supplied example and determine how control points affect the calculated view limits. Done means the axes span the drawn content without being badly squashed, with regression coverage for the example.

Written by the indexing model from the issue text.