Add order option to encoders

[jcrist]

Add order kwarg to encoders

This adds an order kwarg to all encoders for configuring how unordered
collections/objects are encoded. Options are:

• None: the default. All objects are encoded in the most efficient
  manner corresponding to their in-memory representation.
• 'deterministic': Unordered collections (sets, dicts) are sorted
  before encoding. This ensures a consistent output between runs, which
  may be useful when comparing/hashing the encoded binary
  representation.
• 'sorted': same as 'deterministic', but all objet-like objects
  will have their fields encoded in alphabetical order by name. This is
  more expensive than 'deterministic', but may be useful for making
  the output more human readable.

The 'deterministic' output has been heavily optimized - given the work
required to accomplish this feature, I wouldn't expect we can speed up
this operation much more. The 'sorted' option has not been fully
optimized (the assumption being a human-readable output is rarely perf
sensitive). If needed, there are some rather simple optimizations we can
add here to speed this up further.

In general, msgspec.json.encode(obj, order="deterministic") should be
as fast or faster than orjson.dumps(obj, option=orjson.OPT_SORT_KEYS).
For common small object sizes we average a ~20% speedup over orjson
for key sorting.

In [1]: import msgspec, orjson, random

In [2]: enc = msgspec.json.Encoder(order="deterministic")

In [3]: keys = [f'field_{i}' for i in range(6)]

In [4]: random.shuffle(keys)

In [5]: msg = dict(zip(keys, range(len(keys))))

In [6]: %timeit enc.encode(msg)
305 ns ± 2.99 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

In [7]: %timeit orjson.dumps(msg, option=orjson.OPT_SORT_KEYS)
377 ns ± 2.04 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

Fixes #609.