We make sure that negative values are properly compared.

While at it, also make sure that mbstring case conversion takes
into account the specified substitution character and substitution
mode.

Implement full case mapping according to SpecialCasing.txt and
also full case folding according to CaseFolding.txt (F). There
are a number of caveats:

* Only language-agnostic and unconditional full case mapping
  is implemented. The only language-agnostic conditional case
  mapping rule relates to Greek sigma in final position
  (Final_Sigma). Correctly handling this requires both arbitrary
  lookahead and lookbehind, which would require some larger
  changes to how the case mapping is implemented. This is a
  possible future extension.
* The only language-specific handling that is implemented is
  for Turkish dotted/undotted Is, if the ISO-8859-9 encoding
  is used. This matches the previous behavior and makes sure
  that no codepoints not supported by the encoding are
  produced. A future extension would be to also handle the
  Turkish mappings specified by SpecialCasing.txt based on
  the mbfl internal language.
* Full case folding is implemented, but case-insensitive mb_*
  operations continue to use simple case folding. The reason is
  that full case folding of the haystack string may change the
  position at which a match occurred. This would have to be
  mapped back into the position in the original string.
* mb_convert_case() exposes both the full and the simple case
  mapping / folding, where full is the default. The constants
  are:

   * MB_CASE_LOWER (used by mb_strtolower)
   * MB_CASE_UPPER (used by mb_strtolower)
   * MB_CASE_TITLE
   * MB_CASE_FOLD
   * MB_CASE_LOWER_SIMPLE
   * MB_CASE_UPPER_SIMPLE
   * MB_CASE_TITLE_SIMPLE
   * MB_CASE_FOLD_SIMPLE (used by case-insensitive operations)

Instead of using lowercasing.

Instead of performing a binary search, use a hashtable to store
the case maps. In particular a minimal perfect hash construction
is used, which does not require collision resolution (but does
use an auxiliary table for the hash perturbation).

Previously the case mapping table was segregated by the type of
the character (upper, lower, title) and always stored the other
two variants (key, other1, other2). Now the table is segregated
by the target type (key, other). As only very few characters have
more than one target this only slightly increases the size of the
table.

The advantage of this layout is that we only need to perform a
single table lookup in the case table. Previously, depending on
the case that was hit, either one lookup in the property table,
or two lookups in the property table and one lookup in the case
table were required.

This changes the layout from libunicode in the OpenLDAP project
-- however, the last commit there was over 10 years ago, so I
don't see value in keeping this in sync.