Add optimization to sizeof...(X) handling: if none of parameter pack X's

corresponding arguments are unexpanded pack expansions, we can compute the
result without substituting them. This significantly improves the memory usage
and performance of make_integer_sequence implementations that do this kind of
thing:

  using result = integer_sequence<T, Ns ..., sizeof...(Ns) + Ns ...>;

... but note that such an implementation will still perform O(sizeof...(Ns)^2)
work while building the second pack expansion (we just have a somewhat lower
constant now).

In principle we could get this down to linear time by caching whether the
number of expansions of a pack is constant, or checking whether we're within an
alias template before scanning the pack for pack expansions (since that's the
only case in which we do substitutions within a dependent context at the
moment), but this patch doesn't attempt that.


git-svn-id: https://llvm.org/svn/llvm-project/cfe/trunk@284653 91177308-0d34-0410-b5e6-96231b3b80d8