Copy scalars in vector operations to make compiler optimize more

[kronbichler]

If we keep the scalar variables, say a in an update a . x + y, as class variables, the compiler cannot prove that the variable a does not alias with one of the vector arrays. As a result, it keeps re-loading the variable over and over again, even though we know that it won't change. While the compiler can do that for simple code, it does not here because the functors get passed around between multiple functions and call variants, including TBB parallelization. In this PR, it is solved in two different way for the two use cases:

• For the typical vector update loops, we simply create a local copy inside the part that runs the loop on a subrange, which will then signal to the compiler that its value is indeed constant throughout the loop. (Note that the compiler does not rely on the function being marked const, it really needs the local scope.)
• For the reduction operations, we instead copy the functor (and with it, the value of the scalar) as we move into the worker routine. As the functors are small with only some pointer variables, this is cheap enough.

To make sure we do not lose these optimizations and their reason, I added some comments.

[drwells]

Nice.

Could you add your explanation of why some functors are passed by value and others by reference to the header?

[kronbichler]

Could you add your explanation of why some functors are passed by value and others by reference to the header?

I will: In general, the rule is to copy the functors that act on individual elements (reductions), but not the ones still defining loops (vector-add-style functions).