Fix specgram test on NumPy 1.20.

In NumPy 1.20, the fft is slight different (more accurate?) and returns
0 as the last entry, instead of some small number ~1e-30. This causes a
divide-by-zero warning in `log10`. However, this test doesn't really
care about the spectrogram, just the frequencies, so we can change the
scale to linear to avoid the warning.

lib/matplotlib/axes/_axes.py

@@ -7643,6 +7643,7 @@ def specgram(self, x, NFFT=None, Fs=None, Fc=None, detrend=None,
         if mode == 'complex':
             raise ValueError('Cannot plot a complex specgram')
 
+        _api.check_in_list([None, 'default', 'linear', 'dB'], scale=scale)
         if scale is None or scale == 'default':
             if mode in ['angle', 'phase']:
                 scale = 'linear'
@@ -7665,8 +7666,6 @@ def specgram(self, x, NFFT=None, Fs=None, Fc=None, detrend=None,
                 Z = 10. * np.log10(spec)
             else:
                 Z = 20. * np.log10(spec)
-        else:
-            raise ValueError('Unknown scale %s', scale)
 
         Z = np.flipud(Z)