move RNG memory init before particle memory init

[PrometheusPi]

When allocating the RNG memory only 300 MB were free. Thus if the RNG needed a lot of memory (e.g. more than 250MB) no kernel could start. This is avoided by placing the RNG memory allocation before the particle memory allocation.

cc: @Heikman @Flamefire @ax3l

@ax3l This affects the latest release. Please back-port this bug fix.

[PrometheusPi]

@psychocoderHPC I checked this fix without the fix of #1546 - the 250MB of memory are allocated but the simulation does not crash anymore.

[PrometheusPi]

@n01r Then this affects your code as well.

[psychocoderHPC]

Short description without magic numbers:

• one rule in PIConGPU is that it is not allowed to allocate memory after the particles memory pool is created (memory pool allocates all free memory)
• in the current devbranch the RNGProvider is created after the memory pool (not allowed)
  -> this can results in an out of memory error

Solution of this pull request: move the creation of the RNGProvider before the particle memory pool creation