[Hexagon] Remove custom lowering of loads of v4i16

The target-independent lowering works fine, except concatenating 32-bit
words. Add a pattern to generate A2_combinew instead of 64-bit asl/or.

llvm-svn: 308186