TimeDelta misleading 'precision' argument & v4 breaking change

Nobody has claimed this yet.

The docstring is correct:
> :param precision: The time unit used for (de)serialization. Must be one of 'weeks',
> 'days', 'hours', 'minutes', 'seconds', 'milliseconds' or 'microseconds'.

it's the _units_ (de)serialised to; it doesn't effect precision.

That is, 123 microseconds with the default `"seconds"` precision will serialise to `0.000123`, not `0`.

This seems to be a breaking change in v4.0.0, perhaps somehow related to the date/time/isoformat change, I think in 0c901108945573c825e33960707426c80e19290e as previously it would return a `timedelta(microseconds=123)` in the `_serialize` method, but now returns a float via `utils.timedelta_to_microseconds` divided by the 'precision''s mapping value, i.e. `1000000` for seconds.

Start by examining the timedelta serialization path, including the _serialize method and utils.timedelta_to_microseconds referenced in the issue, and reproduce the 123-microsecond example with the default "seconds" setting. Compare the current behavior with the v4 change and determine what the intended precision argument semantics should be; the issue is done when the behavior and compatibility decision are covered by the relevant changes and tests.

Written by the indexing model from the issue text.