This patch does some major cleanup to fix bugs in the mips16 support for 
hardware floating point.  It also adds support for complex return values.  I've 
tested this in a mips32r2 -mabi=32 -mips16 configuration, where it cleans up 
120+ test failures, and also with -mfp64 added to that, where it gets rid of 
~250 failures.  There's also a new test case included to test all the 
combinations of stub functions for interoperability between mips16 and 
non-mips16 mode.

OK to commit?

-Sandra