Improve template argument deduction to keep track of why template

argument deduction failed. For example, given

  template<typename T> struct is_same<T, T> { ... };

template argument deduction will fail for is_same<int, float>, and now
reports enough information

Right now, we don't do anything with this extra information, but it
can be used for informative diagnostics that say, e.g., "template
argument deduction failed because T was deduced to 'int' in one
context and 'float' in another".

llvm-svn: 73237