Cleaned up isWilson/isStaggered/isDwf/getStencilSteps routines

[weinbe2]

This PR addresses issue #1098 and extends it to the getStencilSteps function as well, replacing various functions that look like this:

bool isWilsonType() const
    {
      return (Type() == typeid(DiracWilson).name() || Type() == typeid(DiracWilsonPC).name()
              || Type() == typeid(DiracClover).name() || Type() == typeid(DiracCloverPC).name()
              || Type() == typeid(DiracCloverHasenbuschTwist).name()
              || Type() == typeid(DiracCloverHasenbuschTwistPC).name() || Type() == typeid(DiracTwistedMass).name()
              || Type() == typeid(DiracTwistedMassPC).name() || Type() == typeid(DiracTwistedClover).name()
              || Type() == typeid(DiracTwistedCloverPC).name()) ?
        true :
        false;
    }

And turning them into Dirac abstract functions so each individual class is "forced" to define it so functions like isWilsonType, etc, don't need to be manually updated each time (or, more realistically, don't accidentally get forgotten).

There are some other functions that could this be applied to, hermitian() and hasSpecialMG(), but it seemed reasonable to leave those as false unless proven otherwise.

This PR now replaces those above functions with one-stop shop static Dirac::[...] functions that can query if various QudaDslashType or QudaDiracType enums correspond to Wilson, dwf, or staggered operators. The various Dirac* classes then query these static functions as appropriate. There is scope to apply these functions elsewhere in QUDA.

This PR also fixes a small bug in split grid for staggered vs asqtad type fermions.

[maddyscientist]

I think this description is wrong. It's 1 for operator that consist of a single hopping term, generally full operators, and 2 for composite operators that consist of two hopping terms.

[weinbe2]

It's now true in practice, however there is an exception with the clover hasenbush operator where the "full" operator is still a separate D_eo and D_oe. Getting this counting correct was necessary for the worker abstraction in various operators.

[weinbe2]

I'll update the wording to make that clearer, though

[weinbe2]

83d603f

[maddyscientist]

Having every operator need to specify this seems a bit cumbersome. Why not default all of these operators to return false, and then specialize where needed? DiracWilson can return true and all children will then inherit that. Similarly for staggered and DWF.

[weinbe2]

It was cumbersome for me, but it's a marginal effort for anyone going forward. The reason why I made it abstract and not have it return false by default is to force folks to think about it, but again only once.

I do like the idea of setting it to true for sufficiently high-up "base" classes such as DiracWilson, so I'll pop that in.

[weinbe2]

b4b9e91