Fix key deserialization propagation in windows

[gwaramadze]

Despite key_deserializer being set to json, after windowing, the deserialization is gone, and the key is served back as bytes. Affected windows:

• sliding
• tumbling and hopping in closing strategy "partition" mode

Tumbling and hopping windows with the closing strategy "key" are unaffected.

• The fix for the sliding window was rather trivial. All we need is to use the deserialized key coming directly from process_window instead of the serialized one coming from the store.
  • As a side effect, turns out nothing is utilizing the key from state.expire_windows, so it doesn't need to return it.
• The fix for the "partition" expiration mode is less elegant; refer to the docstring of _deserialize_prefix for a detailed explanation.

[ovv]

Could we store an indicator about the prefix type, if it's a byte or something else ?

For migration, existing windows missing that information can try to deserialize

[gwaramadze]

Yes, but it's considerably more code, not sure if it's worth it.

Let's see what @daniil-quix thinks. But even if we decide to do this, it may be done as a separate optimisation; in the meantime, this will fix the bug.

[ovv]

I'm mostly worried about the performance hit of trying to deserialise every keys

[ghost]

I did some quick tests locally, and the performance hit actually depends on the value itself.

Two almost identical values b'123123' and b'{123123 take very different amounts of time:

### Successful deserialization

%%timeit

try:
    orjson.loads(b'123123')
except Exception:
    ...

34.4 ns ± 0.182 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

### Deserialization failed
%%timeit

try:
    orjson.loads(b'{123123')
except Exception:
    ...

765 ns ± 4.51 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

[ghost]

This was kinda testing that the same prefix is returned :)
Let's return the key from expire_windows() to make the tests legit, even though we don't use it in the actual code.