Change compound assignment operators to keep track of both the promoted

LHS type and the computation result type; this encodes information into 
the AST which is otherwise non-obvious.  Fix Sema to always come up with the 
right answer for both of these types.  Fix IRGen and the analyzer to 
account for these changes.  This fixes PR2601.  The approach is inspired 
by PR2601 comment 2.

Note that this changes real *= complex in CodeGen from a silent 
miscompilation to an explicit error.

I'm not really sure that the analyzer changes are correct, or how to 
test them... someone more familiar with the analyzer should check those 
changes.

llvm-svn: 67889