This is, for the majority of backends, the native format anyway, and so
we save encoding and decoding when using the PickleCodec or if (in future)
a orjson Coder was to be added.
For the JsonCodec, the only thing that changed is the location where the
JSON data is encoded to bytes and decoded back again to a string.
In `timeit` tests, 10.000 calls to `ModelField()` could take up to half
a second on my Macbook Pro M1, depending on the type annotation used.
Given that the method is called for every cache hit, this can really add
up. The number of different return types for endpoints is very much
finite however, so caching is a definite win here.
Use the return annotation to decode cached data to the correct type.
This follows the same logic FastAPI uses to JSON request bodies.
For the PickleCoder, this is a no-op as pickle already stores type
information in the serialised data.
