More accurate prognosis for CPU memory

[myurkin]

Currently OpenCL has very convenient algorithm for processing -prognosis, which counts memory inside the allocation functions (seamlessly for the main code). We should try to do the same thing for the CPU memory. The problem with the latter is that some memory (i.e. in make_particle.c) is always allocated. We may circumvent this by passing additional arguments to allocating (and maybe freeing) functions, which will be either prognosis or true depending on the type of the array. It would also be great if the freeing part can be handled fully automatically.

[myurkin]

Moreover, the execution flow on the CPU is heavily affected by prognosis, so it may be hard, e.g., to call allocation inside the disabled block.

[myurkin]

This is relevant not only for prognosis, but for CPU-memory calculation in general

[myurkin]

An alternative approach is to use OS-dependent calls to obtain peak memory. This will obviously be less informative in prognosis mode.
Windows - GetProcessMemoryInfo()
Linux/MacOS - getrusage()