We've historically expected that the return values from libmagic are
ascii, since they are constant strings or stuff like dates/numbers.
In some cases, however, it will return information like the title of
the document in the doc's native character set, which is unknown to
us.  This produces decode errors.

I have not been able to get a document that triggers this behavior,
but the safest change is to decode with 'backslashreplace', which
escapes undecodable characters with a backslash.

extension support was added to libmagic in 2015 in
9190a18d09f25fb0ca6abe1fcbdba780f5077e45.  This change:

 - updates the test to handle various verions that return garbage /
   empty string.s
- throws an exception in magic.Magic if extension support is requested
  and we know its not supported.

Some versions of libmagic (unclear which) fail the setparam call in
the Magic constructor.  This this is just resolving a corner case to
begin with, silently continue on if this happens.

Potentially `magic_open` may fail, and when python attempts to gc the Magic instance it fails with:
```
Exception ignored in: <function Magic.__del__ at 0x7f595ba06430>
Traceback (most recent call last):
  File "/usr/local/lib/python3.8/site-packages/magic.py", line 129, in __del__
    if self.cookie and magic_close:
AttributeError: 'Magic' object has no attribute 'cookie'
```

These aren't present in ancient versions of libmagic, so only fetch
optionally and throw NotImplementedError if called.

Not going to change the value of MAGIC_MIME until a major version
bump, but at least add the missing MAGIC_MIME_TYPE constant and use
where it should be.

https://github.com/ahupp/python-magic/pull/70