Don't redefine str(n)casecmp on windows unless they're missing

[nmathewson]

When we do redefine them, use inline functions instead of #define.

This fixes a latent code problem in our redefinition of these
functions, which was exposed by our refactoring: Previously, we
would #define strcasecmp after string.h was included, so nothing bad
would happen. But when we refactored, we would sometimes #define it
first, which was a problem on mingw, whose headers contain
(approximately):

inline int strcasecmp (const char *a, const char *b)
{ return _stricmp(a,b); }

Our define turned this into:
inline int _stricmp(const char *a, const char *b)
{ return _stricmp(a,b); }

And GCC would correctly infer that this function would loop forever,
rather than actually comparing anything. This caused bug 26594.

Fixes bug 26594; bug not in any released version of Tor.

[coveralls]

Coverage decreased (-0.02%) to 59.869% when pulling c8ccd02 on nmathewson:bug26594 into cb1a367 on torproject:master.

[gvanem]

No, no:

static inline int strncasecmp(const char *a, const char *b, size_t n) {
  return strnicmp(a,b,n);
}
...
static inline int strcasecmp(const char *a, const char *b) {
  return stricmp(a,b);
}

[nmathewson]

Thanks! I'll use the variant with an underscore instead, though, since msdn says that the unprefixed versions are deprecated. Tried to fix this with fecb821.

[gvanem]

Good. Thanks!