As far as I can tell it is simply wrong per-RFC to put a charset parameter on a mulitpart content-type.  So I think this should, indeed, raise an error on the Multipart subtype.

If someone sets any charset, the CTE is set wrong.  So code that sets charset is already broken, even though tolerant mailers will accept the resulting message (but some mailers won't, as described in this issue).

So, I think set_charset on MIMEMultipart should be deprecated and turned into a no-op in 3.2.