Annarchy

* Bugfixes:
* In rare cases the 'w' variable could not be recorded.
* Bugfix in early stopping of simulations (simulate_until).
* Bugfix in CUDA code generation using dense matrix format.
* Bugfix in code generation related to Runge-Kutta 4th order method and refractoriness.
* Bugfix in cython wrapper for Monitor (github issue 3)

* `reset()` has a new argument *monitors* to control the clearing of monitored results.
* Improved error messages for non-implemented features on CUDA devices.
* Bugfixes:
* fixed bugs related to compuation of mean-firing rate for spiking models.
* fixed a bug related to IndividualNeuron accessor.

* Extended the available set of hand-written SpMV codes (in particular SSE and AVX-512)
* Bugfixes:
* the usage of storage_format="dense" and a constant weight (e.g. in the COBA example) leads to an error message.
* Network.enable_learning() had no period/offset arguments.
* TypeError occured in connect_fixed_number_post() cause by a type mismatch on a temporary variable.

* Bugfixes:
* Handling of non-uniform delays in connect_with_func
* Added Monitor.reset() to reinitialize monitors. Called automatically by reset().

* Changed default compiler flags used by ANNarchy:
* cython modules build during setup.py use only `-O3`
* c++ (model compile) default flags are now: `-O3 -ffast-math -fno-finite-math-only -march=native` (following Brian2 default flags)
* Improved automatic format selection for rate-coded models (experimental, we greatly appreciate bug reports).
* HomogeneousCorrelatedSpikeTrains: `corr` can now be a single value.
* TimedPoissonPopulation: `rates` can now be a single value.

* Bugfixes:
* Monitor: fixed start and stop times.
* CUDA: device configuration (annarchy.json) was ignored.

* Automatic format selection for rate-coded models (experimental, we greatly appreciate bug reports).
* Bugfixes for data structures introduced in 4.7.0.
* Improved code generation, e. g. AVX instructions for default psp of rate-coded models, added dense matrices for spiking models.