Since the dawn of time, filesort has had different paths for making sort keys
for fields and items. There's no good reason to keep carry this distinction
forward anymore, so remove the special handling of Field and just sort Items
everywhere, making filesort somewhat simpler. There are some special legacy
cases with sorting enums etc. where we still need to look into Item_field and
look at the actual underlying file type, but they are fairly localized and now
much more visible.

There are some edge cases where keys used to be shorter, e.g. all integers
will now get 64-bit sort keys instead of 8-bit, 16-bit, etc.. Also, some
EXPLAIN output changes, as we also no longer have specialized formatting of
Field.

WEIGHT_STRING() now is able to get the actual weight strings for more kinds
of items, in particular integers.

Inadvertently, we fix a bug that was only half-visible in the tests;
common-tests.inc contain a query whose results were not properly deduplicated
before, but are now, and while this was hidden by LIMIT/OFFSET, it affects
which rows are visible in the result set.

sysbench sorting tests (ranges and distinct ranges) show strong improvements
at high thread counts; as much as +17%. Lower thread counts are neutral or
in the noise.

Unfortunately, replication started using Field::make_sort_key() in 8.0, and
cannot easily change it back, as they need exact byte-by-byte matches of the
data. Fixing this would enable removing 500+ lines of complicated sort key
code.

Change-Id: Ieb5038e8afbea19a5f0e0e7f83b3177cb3176e84