Check the linkage cache at every recursive step.

Before this patch the linkage cache was only used by the entry level function
(getLinkage). The function that does the actual computation (getLVForDecl),
never looked at it.

This means that we would not reuse an entry in the cache when getLVForDecl did
a recursive call. This patch fixes that by adding another computation enum
value for when we don't care about the linkage at all and having getLVForDecl
check the cache in that case.

When running "clang -cc1" over SemaExpr.ii this brings the number of linkage
computations from 93749 to 58426. When running "clang -cc1 -emit-llvm -O3" it
goes from 198708 to 161444.

For SemaExpr.ii at least linkage computation is a small enough percentage of
the work that the time difference was in the noise.

When asserts are enabled this patch also causes clang to check the linkage
cache even on recursive calls.

llvm-svn: 182799