Replace Ghostscript DOCINFO and fix 9.25 metadata date regression

We no longer use Ghostscript to manage PDF metadata, instead
omitting the DOCINFO segment from the pdfmark file we generate.

Instead all of the relevant metadata code has been migrated to pikepdf,
and we use that API. This should be more consistent and fixes the
Ghostscript version-depedent quirks.

Also removes our python-xmp-toolkit dependency, except for
testing.

setup.py

@@ -255,7 +255,6 @@ setup(
         'Pillow >= 4.0.0, != 5.1.0 ; sys_platform == "darwin"',
                                   # Pillow < 4 has BytesIO/TIFF bug w/img2pdf 0.2.3
                                   # block 5.1.0, broken wheels
-        'python-xmp-toolkit >= 2, < 3',
         'reportlab >= 3.3.0',     # oldest released version with sane image handling
         'ruffus >= 2.7.0',
     ],