Sparse Core: Replace malloc/free with conditional_aligned

Reference issue

What does this implement/fix?

The sparse classes currently use a mix of std::malloc, std::realloc, and std::free for memory management instead of the aligned_malloc family of functions defined in Memory.h. Other than consistency with the dense classes, this will enable users to track heap allocations with #define EIGEN_RUNTIME_NO_MALLOC and related mechanisms. Also, there may be some advantage to ensuring that the sparse matrices use aligned storage for vectorization.

Additional information

Benchmark measurement:

double sse no align Total duration: 21531, per iteration: 2153.1
double sse align    Total duration: 21691, per iteration: 2169.1
double avx no align Total duration: 21548, per iteration: 2154.8
double avx align    Total duration: 21448, per iteration: 2144.8
double old sse      Total duration: 95201, per iteration: 9520.1
double old avx      Total duration: 97138, per iteration: 9713.8

I did this test 10 times for size = 1000.

template<typename Scalar, bool Align>
void testSparseAlign(Index size)
{
  SparseMatrix<Scalar, ColMajor, int, Align> A(size, size);
  for (int i = 0; i < size; i++)
    for (int j = 0; j <= i; j++)
      A.coeffRef(size-i-1, j) = i*j;

  std::cout << A.nonZeros() << "\n";
  std::cout << A.data().allocatedSize() << "\n";
  std::cout << A.sum() << "\n";
}