rewrite diag code in c

[qzhu2017]

For a large structure, diag() function becomes slow.

[qzhu2017]

[qzhu2017]

Next step is to MPI on Kef_C

[qzhu2017]

@omioisaac23
I think the new code is much faster now,
1 week ago, a single predict structure call take 90 seconds
after the MPI of diag/descriptor calculation, it was 35 seconds
after the MPI of K_ee/K_ef, it is only 7 seconds !

[qzhu2017]

CPU-16

CPU-32