When we suppress an error due to SFINAE, stash the diagnostic away with the

overload candidate, and include its message in any subsequent 'candidate not
viable due to substitution failure' note we may produce.

To keep the note small (since the 'overload resolution failed' diagnostics are
often already very verbose), the text of the SFINAE diagnostic is included as
part of the text of the note, and any notes which were attached to it are
discarded.

There happened to be spare space in OverloadCandidate into which a
PartialDiagnosticAt could be squeezed, and this patch goes to lengths to avoid
unnecessary PartialDiagnostic copies, resulting in no slowdown that I could
measure. (Removal in passing of some PartialDiagnostic copies has resulted in a
slightly smaller clang binary overall.) Even on a torture test, I was unable to
measure a memory increase of above 0.2%.

llvm-svn: 156297