- endianness issues: on big endian machines, writing out only part
   of an integer was broken (get_needed_bytes(x) < sizeof(x))
    -> shift integer before memcpy() on big endian machines

 - alignment issues: unaligned reads when loading binary plist. Leads
   to slow runtime performance (kernel trapping and fixing things up),
   SIGBUS (kernel not helping us out)
    -> introduce get_unaligned() and have the compiler generate the code
       needed for the unaligned access
   (note that there remains unaligned accesses that I haven't been able
    to track down - I've seen 2 of them with test #2)

 - type-punning issues: breaking strict aliasing rules can lead to
   unexpected results as the compiler takes full advantage of the aliasing
   while optimizing
    -> introduce the plist_uint_ptr union instead of casting pointers

Tested on amd64, alpha and hppa.

In locales like German, a ',' is used as a decimal separator. When the
program calling plist_to_xml uses LC_NUMBER with something different
than a '.', parsing of the resulting XML document fails. This patch
fixes it.

This makes it possible to process the resulting char* directly as
a c-string without further copying.

* on armel system floating poing data can have different endianess than
  rest of types; hence we fix arm endianess for defined(__VFP_FP__) to
  be big/native; this also applies for data parsing/writing
* date parsing didnt flip the endianess back for little endian systems
  when reading the values causing test failures; we fix this by ensuring
  float endianess is applied when parsing