Even though we are destructively changing the recycled cr_Packages data,
we can keep the structure self-contained so the upper level logic will
automatically take care of the allocated memory.

The recycled metadata cr_Packages use one chunk per all packages
(CR_PACKAGE_SINGLE_CHUNK).  That's why (pkg->chunk == NULL) and we can
re-use it for a new chunk.  We just need to change the type of pkg to
!CR_PACKAGE_SINGLE_CHUNK so the cr_package_free() cleans the new
chunk.

This allows us to stop taking care of the locally allocated location_*
variables, because they are _cleanup_free_.  This makes even the buffer
handling simpler.

Revamped 'pkg_from_md' that was unused for some time, but renamed to
'clean'.

The 'task_cleanup' target was reached without free-ing the 'pkg'
structure in one of the error cases before.

The 'res' was allocated on two places.

Will be needed later for the "skip" mechanism.

Normally we dump (to XML and DB) directly from the pool of threading
workers, in a streaming fashion (as soon as the metadata about packages
are determined it is written).

The new option causes that the metadata are loaded first (still in
parallel), and then written at one time (by one thread for now).
Since all the metadata (cr_Package) is available in the memory in one
time, we can do some additional processing and filtering (before we
eventually perform the dump).

This is an optional behavior because the delayed XML chunk preparation
slows down the overall multi-threaded processing time.

Commit 8fb99cd added a detection
mechanism for duplicated NEVRA packages and an useful warning.

The new option '--duplicated-nevra keep' keeps the default behavior (if
specified) — all the duplicate packages are dumped into metadata.

When '--duplicated-nevra keep-last' argument is used, only the last
NEVRA (ordered by build time, then by its ocation path) is dumped to
the metadata and the former duplicates are skipped.

Relates: https://pagure.io/copr/copr/issue/840
Fixes: 325