This does not actualy fit the project name "liblongpath", but most of the
surrounding code is identical. So for the time being I will tolerate
the fact that liblongpath contains a little side use, but should we want
to integrate more functionality, we should probably give the project a
more generic name.

Use of the interposer by environment variables

LD_PRELOAD=<path_to_lib>

* standard behaviour: map O_DIRECT to O_DSYNC for open() failing
  with EINVAL
* set CONVODIRECT=none to just ignore O_DIRECT

(who the hell uses these legacy calls? java doess...)

bugfix: With slashes near a multiples of PATH_MAX - 1 into the path, liblongpath could return results for a wrong path

skip all trailing slashes before starting the search, this is
required because otherwise the 'rest' path could start with a
slash (would work on wrong path) or be the empty string
(yield undefined error)

(unless someone tells me how to do this better (in any practical way))

platform dependencies, remove functions we invented, done implement functions which don't exist on the platform

linux needs sys/time.h
no lutimes on solaris
also fix return value of readlinkat (also _s_size_t)
there is no rmdirat, remove our invention
only implement mkostemp* where they exist

our() macros don't work for our internal _openat, fix readlink type and struct stat, more interpose prep

have checked the cpp output, all it does is shuffle around the definitions:

for f in src/*.i ; do egrep ^extern /tmp/liblongpath/$f | sort >/tmp/liblongpath/$f.s && egrep ^extern $f | sort >$f.s && diff -ub /tmp/liblongpath/$f.s $f.s|less ; done

this commit does _not_ change cpp (ggc -E) output compared to cdb63038
(tested with the commands given in src/Makefile.am for Solaris and Linux)

better indentation for include files