GitLab

So there is this BlockProcessing.h that fortunately noone uses, which
expects a compile time NTHREADS for some reason. Added a note for
whomever gets burned with a compilation error.

Not sure about this, but now Cubism automatically defines
`CUBISM_ALIGNMENT` if it is not already defined.
That means that if the app intends to define the value,
it must define it before including _any_ Cubism header.

An alternative would be to put these ifdefs to Matrix3D.h (the only file
using the constant), but probably better to be strict.

C++ reserves all names starting with `_X` for some capital letter `X`.
Instead, use `CUBISM_` prefix.
Cubism still accepts all macros, but pops out a warning if used.

Instead of using _BLOCKSIZE*_ macros, use Block::sizeX/Y/Z.
Slice dumper currently is not implemented for non-cubical blocks.
pack/unpack functions get block size as a runtime argument, which
shouldn't really affect performance.

In tests, avoid _Capitalized, as that formatting is reserved by C++.

Removing _HDF5_DOUBLE_PRECISION_ from the library.

It was used only by the old coupling framework examples.
Now we require the user app to specify the Block class by itself.

User app must specify `Block::Real` and `ElementType::Real`.

Implemented using a #define trick I saw in pybind11.
If necessary we can a) change the namespace name, b) get rid of defines.

Wrong file offsets were aliasing memory addresses which cause corrupted
data and partial serialization of file write operations.

template<typename TGrid, typename TStreamer> should be
template<typename TStreamer, typename TGrid>

They are defined in their own namespaces.  More general this way.