Previously, GitPython chokes on this while decoding.  Rather than
choking, instead accept the error and replace the invalid bytes by the
� (\x80) char.

This adds 2 properties to commits.  Their values are derived from the
existing data stored on them, but this makes them more conveniently
queryable:

    - authored_datetime
    - committed_datetime

These return "aware" datetimes, so they are effectively companions to
their raw timestamp equivalents, respectively `authored_date` and
`committed_date`.

These datetime instances are convenient structures since they show the
author-local commit date and their UTC offset.

The problem is that a per-tree modification API cannot work properly, as the sorting is based
on full paths of all entries within the repository. This feat can only be achieved by the index,
which to my knowledge already does it correctly.

The only fix is to remove the misleading API entirely, which will happen in the next commit.

Related to #369

[skip ci]

Previously it was possible to generate trees which didn't
appear legit to git as gitpython's sorting was a simple alpha-numeric
sort. Git uses one that minimizes literal string comparisons though,
and thus behaves slightly differently sometimes.

Fixes #369

Related to #362

Previously timezones which were not divisable by 3600s would be
parsed correctly, but would serialize into a full hour, rounded up.

Now floating point computation is used which fixes the issue.

Related to #336

Without this commit the update() function of a submodule does not work
in CentOS 6.