Support optionally using leptonica to deskew

unpaper doesn't seem to be good at deskewing. It fails on test case
with a lot of italics. I think it also struggles on pages with a lot
of whitespace. Leptonica continues to shine here.

However, this is only a first crack at Leptonica. The leptonica module
should be redone to use cffi (more extensible).

Also considering the possibility of making all Lept calls in a forked
process to insulate the calling process from C code crashes and the
messy redirect of stdout/stderr to read Leptonica's errors.

I don't think the redirect is a huge problem as long as multiprocesses
rather than multithreads are used. The ruffus child process that is
handling a page is single threaded and will not be affected by the
redirection. It just feels dirty. The main reason to consider a child
process is crash isolation.

ocrmypdf/main.py

@@ -185,6 +185,9 @@ advanced.add_argument(
     '--tesseract-timeout', default=180.0, type=float, metavar='SECONDS',
     help='give up on OCR after the timeout, but copy the preprocessed page '
          'into the final output')
+advanced.add_argument(
+    '--deskewer', choices=['leptonica', 'unpaper'], default='leptonica',
+    help='choose deskew provider')
 
 debugging = parser.add_argument_group(
     "Debugging",
@@ -493,7 +496,11 @@ def preprocess_deskew(
     pageinfo = get_pageinfo(input_file, pdfinfo, pdfinfo_lock)
     dpi = int(pageinfo['xres'])
 
-    unpaper.deskew(input_file, output_file, dpi, log)
+    if options.deskewer == 'unpaper':
+        unpaper.deskew(input_file, output_file, dpi, log)
+    else:
+        from . import leptonica
+        leptonica.deskew(input_file, output_file, dpi)
 
 
 @transform(