Make SPARC hardware sqrt/f functions compatible with Clang/LLVM

Compiling the current libmcs with hardware sqrt enabled for SPARC using Clang/LLVM leads to the following errors:

sqrtd.c:64:32: invalid output constraint '=f' in asm
sqrtf.c:11:32: invalid output constraint '=f' in asm

Update the inline assembly to support also Clang/LLVM. Also make sure that any extended inline assembly does not cause a performance penalty when using GCC.