Delay implementation kernel cleanup

[JanVogelsang]

PR #2989 adds axonal delays to NEST, but still leaves some open ToDos for a follow-up PR. Such a follow-up PR should clean up existing suboptimal design choices made in NEST a long time ago, which were made even more pronounced by PR #2989. This includes:

• Get rid of special treatment for delays (and weights maybe) when setting synapse params
• Move all sanity checks for delay combinations in a single location (instead of CopyModel/SetDefaults, synapse params, SetStatus)

Furthermore, PR #2989 only adds axonal delay support to a single synapse (stdp_pl_synapse_hom_ax_delay) and neuron model (iaf_psc_alpha) so far. Most other plastic synapse models could also benefit from axonal delays, but not all users will actually require them. To provide the best possible performance for each of both cases, there should be different implementations for each case. This can be done elegantly by still using the same base synapse model implementation for both total-delay and axonal-dendritic-delay containers (which represent the both cases, see delay_types.h) and optimize only parts the code (e.g. send function) using compile-time directives (e.g. if constexpr or overloading the send function based on delay implementation).