Eliminate a silly little Parse/Sema dance when parsing typename

Eliminate a silly little Parse/Sema dance when parsing typename
specifiers such as

  typename T::template apply<U>

Previously, we would turn T::template apply<U> into a
TemplateSpecializationType. Then, we'd reprocess that
TemplateSpecializationType and turn it into either a
TemplateSpecializationType wrapped in an ElaboratedType (when we could
resolve "apply" to a template declaration) or a
DependentTemplateSpecializationType. We now produce the same ASTs but
without generating the intermediate TemplateSpecializationType.

The end goal here is to avoid generating TemplateSpecializationTypes
with dependent template-names, ever. We're not there yet.

git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@126589 91177308-0d34-0410-b5e6-96231b3b80d8