Fix for Ifpack2 that causes SPARC slowness.

Created by: kyungjoo-kim

Description

Last a few commits that I made for Ifpack2 cause slowness in SPARC. The reason for the slowness is surprisingly from some template instanciations. I cannot precisely explain why. Previously we use macros for blockwise specialization and I changed it to template implementation for doing the same thing. After changing it back to macro expressions, it recovers the previous performance. Sigh Sigh Sigh.... Mystery....

How Has This Been Tested?

This is tested on Apollo and @vbrunini confirmed that it is fixed on SPARC.