Document unpickle on scheduler changes

[crusaderky]

This paragraph has become obsolete after #7564 and it's already caused confusion: https://dask.discourse.group/t/scheduler-importing-modules-meant-for-the-workers/2281

distributed/docs/source/protocol.rst

Lines 121 to 151 in 7d0ba9c

	Cross Language Specialization
	-----------------------------
	The Client and Workers must agree on a language-specific serialization format.
	In the standard ``dask.distributed`` client and worker objects this ends up
	being the following::
	bytes = cloudpickle.dumps(obj, protocol=pickle.HIGHEST_PROTOCOL)
	obj = cloudpickle.loads(bytes)
	This varies between Python 2 and 3 and so your client and workers must match
	their Python versions and software environments.
	However, the Scheduler never uses the language-specific serialization and
	instead only deals with MsgPack. If the client sends a pickled function up to
	the scheduler the scheduler will not unpack function but will instead keep it
	as bytes. Eventually those bytes will be sent to a worker, which will then
	unpack the bytes into a proper Python function. Because the Scheduler never
	unpacks language-specific serialized bytes it may be in a different language.
	**The client and workers must share the same language and software environment,
	the scheduler may differ.**
	This has a few advantages:
	1. The Scheduler is protected from unpickling unsafe code
	2. We could conceivably implement workers and clients for other languages
	(like R or Julia) and reuse the Python scheduler. The worker and client
	code is fairly simple and much easier to reimplement than the scheduler,
	which is complex.
	3. The scheduler might some day be rewritten in more heavily optimized C or Go

CC @fjetter