Array wrapping enhancement

[sshiraiwa]

This PR includes

1. address recent addition of kernel_dispatch.hpp in MFEM C++.

• kernel_dispatch.hpp doesn't need to be exposed to Python. But, we need MFEM_REGISTER_KERNELS macro to be
  defined somehow. In this PR, MFEM_REGISTER_KERNELS is replaced to be empty macro. So that SWIG does not
  see the lines containing this macro.

2. Add CoefficentArray, VectorCoefficientArray, and MatrixCoefficientArray (represent Array<* Coefficient> and so on)

• Changed Array constructor behavior when list of SWIG-wrapped MFEM objects is passed. SWIG objects passed as an item in list/tuple looses owner ship to underlying C++ MFEM object. Internally, we create a new Array (depending on the length of container) and set each Array element one by one (after setting thisown to False).

3. Renamed other Array< * MFEM object>.

• They are used to be named like VectorPtrArray. We simplify it as VectorArray.

4. SparseMatrixArray2D, DenseMatrixArray2D, HypreParMatrixArray2D are added

• HypreParMatrixArray2D can be used with mfem.HypreParMatrixFromBlocks
• DenseMatrixArray2D appears in BlockNonlinearFormIntegrator
• SparseMatrixArray2D is used wtih MixedBilinearForm::GetBlocks

ToDo:

• Update Doc
• Add test/test_array2d.py

[sshiraiwa]

PR is made to trigger the test.

[sshiraiwa]

Example asked in #253 can be done as follows.

   tensors =  [mfem.DenseMatrix(np.ones((3,3))*i) for i in range(max_attr)]
   tensor_coefs =  [mfem.MatrixConstantCoefficient(mat) for mat in tensors]
   sigma_array = mfem.MatrixCoefficientArray(tensor_coefs)
   sigma_attr = mfem.intArray([1,2,3,4,5])
   sigmaCoef = mfem.PWMatrixCoefficient(dim, sigma_attr, sigma_array, False)

[justinlaughlin]

@sshiraiwa let me know when this is ready I'm happy to test it out.

[sshiraiwa]

Closing this one w/o merging, since this version will be baseline of PyMFEM4.8