Use an array-of-scalars as the base case for Tensor.

[bangerth]

In #16474, I replaced the data structure to store the elements of a tensor from a C-style array to a std::array. But for rank-1 tensors, this is still an array of Tensor<0,dim>. This patch breaks the recursion one level higher: Rather than going to to std::array<Tensor<0,dim>,dim>, we now use std::array<Number,dim>.

The marginally non-trivial part is that we now have to initialize this array where before the default constructor of Tensor<0,dim> did that. This is awkward because std::array by default does not initialize its elements if they are built-in types. The first commit lets std::array not initialize its elements, and then uses std::fill to initialize the elements.

The second commit avoids this two-step process by direct initialization of the elements using a brace-enclosed initializer list. This took me a little bit to come up, and I will admit that the solution is perhaps not easy to understand at first read. But it is localized, and I tried to document it well enough. It uses some fun features of C++ :-)

Related to #16465.

[bangerth]

The third commit also adjusts value_type and array_type accordingly.

Following commits work around compiler warnings :-(

[kronbichler]

Thank you @bangerth for the explanations; I am fine with this version (albeit I would probably been too lazy myself to go away from fill).

[kronbichler]

Would you mind squashing together some of the commits that do not work on their own. (I realize it might not be worth to restructure the commits cleanly, so feel free to ignore this request.)

[bangerth]

I made the commits more self-contained and complete.