Commit.diff used to be more than happy to accept any Unicode or byte

strings thrown at it, and some code actually expected these to be

Unicode strings, attempting to unconditionally encode the contents as a

UTF-8 byte string. If a hosting service set the diff as a byte string,

and there was Unicode content within the diff, this could lead to a

crash.
Now, Commit.diff always stores a byte string, handling the encoding

from Unicode if needed. Callers can and should now always treat this as

a byte string.
This fixes a crash when posting existing commits from Bitbucket with

Unicode content.

Unit tests pass.
Manually tested that the diff content no longer gets improperly

re-encoded and crashes in the customer case.