Refactor serialization mechanism to be extendable

At the moment, our serialization code is just a really long `if-elif-elif-...` block: https://github.com/elastic/apm-agent-python/blob/ee45d21ff43be495089a005863b7f6af0707fe34/elasticapm/utils/encoding.py#L85-L138.

This has several drawbacks:
* it's not extendable (we e.g. might want to serialize Django `QuerySet` objects in a sane way that doesn't evaluate the queryset unnecessarily, but only if Django is installed)
* it's brittle
* it's hard to read
* ideally, serialization would also include shortening of values, which is hard to do in its current state

One option could be to have a registry, to which transformers can register themselves, together with a list of supported types for that transformer.

Start by reading elasticapm/utils/encoding.py around lines 85-138, where serialization is currently implemented as a long conditional chain. Compare the proposed registry approach with the need for optional Django QuerySet handling and value shortening. Done means the serialization mechanism is extensible, less brittle, and avoids unnecessarily evaluating QuerySets when Django is installed.

Written by the indexing model from the issue text.