Examine the preserve_nan method in replace_epochs

[bburan]

I'm skeptical this is required. This suggests there's a deeper bug. For example, this works just fine without needing the preserve NaN magic:

x = np.arange(10, dtype=np.float)
y = np.arange(2, dtype=np.float)
y[0] = np.nan

print(x)
print(y)

x[4:6] = y
print(x)