Refactor the Multipart parsing into a Sans-IO layer

This allows it to be used in async (ASGI) frameworks. It also
hopefully makes the code a little clearer to follow.

This removes the ``Content-Transfer-Encoding`` support as RFC7578
states it is deprecated and

    Currently, no deployed implementations that send such bodies have
    been discovered.

This requires the dataclasses backport, which I think is ok as
dataclasses are in the stdlib, and 3.6 has less than a year till EoL.

The API is based on that successfully used by h11.

docs/http.rst

@@ -158,5 +158,3 @@ environments for unittesting you might want to use the
 .. autoclass:: FormDataParser
 
 .. autofunction:: parse_form_data
-
-.. autofunction:: parse_multipart_headers

setup.py

@@ -1,4 +1,8 @@
 from setuptools import setup
 
 # Metadata goes in setup.cfg. These are here for GitHub's dependency graph.
-setup(name="Werkzeug", extras_require={"watchdog": ["watchdog"]})
+setup(
+    name="Werkzeug",
+    install_requires=["dataclasses; python_version < '3.7'"],
+    extras_require={"watchdog": ["watchdog"]},
+)

src/werkzeug/_internal.py

@@ -131,7 +131,7 @@ def _to_str(x, charset=_default_encoding, errors="strict", allow_none_charset=Fa
     if x is None or isinstance(x, str):
         return x
 
-    if not isinstance(x, bytes):
+    if not isinstance(x, (bytes, bytearray)):
         return str(x)
 
     if charset is None: