Don't leak TemplateIds when a plugin parses late-parsed templates at TU end.

In -fdelayed-template-parsing mode, templates that aren't used are not parsed
at all.  For some diagnostic plugins, this is a problem since they want to
analyse the contents of the template function body.  What has been suggested
on cfe-dev [1] is to explicitly parse interesting templates in
HandleTranslationUnit(); IWYU does this for example [2].

This is workable, but since the delayed parsing doesn't run below a call to
ParseTopLevelDecl(), no DestroyTemplateIdAnnotationsRAIIObj object is on the
stack to clean up TemplateIds that are created during parsing.  To fix this,
let ~Parser() clean them up in delayed template parsing mode instead of
leaking (or asserting in +Assert builds).

(r219810, relanded in r220400, fixed the same problem in incremental processing
mode; the review thread of r219810 has a good discussion of the problem.)

To test this, give the PrintFunctionNames plugin a flag to force parsing
of a template and add a test that uses it in -fdelayed-template-parsing mode.
Without the Parser.cpp change, that test asserts.

1: http://lists.cs.uiuc.edu/pipermail/cfe-dev/2014-August/038415.html
2: https://code.google.com/p/include-what-you-use/source/detail?r=566 


git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@237531 91177308-0d34-0410-b5e6-96231b3b80d8