Kernel fusion for 4-d preconditioned domain-wall and Mobius kernels

[maddyscientist]

The 4-d preconditioned chiral fermion dslashes presently use multiple kernels to apply the preconditioned dslash. This is an issue to explore fusing these kernels, the motivation is two fold:

1. This will improve single node performance since it will reduce memory traffic.
2. If the 4-d dslash part is fused with other operations it will make the strong scaling better since there will be more compute to overlap with.

[weinbe2]

Mobius is covered by https://github.com/lattice/quda/pull/1163 , I guess 4-d preconditioned Shamir is still outstanding