Document the synchronisation behavious of alpaka buffers

[fwyzard]

As far as I know, we do not document any synchronisation behaviours for alpaka buffers.
However, different buffers and different back-ends effectively implement different behaviours.

• A buffer allocated on a CUDA gpu with alpaka::allocBuf() internally used cudaMalloc() to allocate the memory, and cudaFree() to release it. While not mentioned explicitly in the CUDA documentation, the observed behaviour is that cudaMalloc() and cudaFree()¹ are blocking and synchronise across all kernels executing on the current GPU.

• A buffer allocated on the host with alpaka::allocMappedBuf() for the CUDA platform internally uses cudaMallocHost() and cudaFreeHost(). I have no idea if these are blocking calls that synchronise with the current (or all) CUDA device(s).

• A buffer allocated on the host with alpaka::allocBuf() internally uses aligned new and delete. These do not imply any synchronisations.

• A buffer allocated on the host with alpaka::allocMappedBuf() for the CPU platform uses allocBuf(), and does not imply any synchronisations.

• A buffer allocated with alpaka::allocAsyncBuf() should be queue-ordered and asynchronouse with respect to the host. But we should double check that it is indeed the current behaviour :-)

--
¹ except for memory allocated by cudaMallocAsync().

[fwyzard]

Should we add new functions that implement a common behaviour across all backends ?

[fwyzard]

See also #2417 .