Improve (de-)serialization performance for scalar arrays

[124C41p]

Fixes #515

Since (de-)serialization is implemented purely in Python, it is quite slow compared to native implementations. I try to circumvent that issue by not deserializing repeated scalar fields immediately, but wrapping their byte representation inside the ScalarArray[T] class instead. This class acts like a list. That is, you can call len(a), a[i], and list(a) for any ScalarArray a, and only at this point we actually deserialize (which is still very slow for big arrays).

On the other hand, when using numpy you can also call np.asarray(a) for any ScalarArray a to turn it into a numpy array in no time. Conversely, any numpy array b can be turned into a ScalarArray by calling ScalarArray.from_numpy(b) to be passed to a betterproto dataclass field (instead of a list) for faster serialization speed.

I tried to be as non-breaking as possible. That is, you can use lists everywhere you used them before. However, it was necessary to generate Sequence[T] type hints where List[T] hints were generated before. Also note that ScalarArray is an immutable data structure. So you might not be able to use .append() or .insert() on repeated fields as before (although it should be possible to make ScalarArray mutable if really needed).

What do you think about this approach?

[Gobot1234]

Superseded by #545