genn-team/genn: GeNN 3.2.0

This release extends the initialisation system introduced in 3.1.0 to support the initialisation of sparse synaptic connectivity, adds support for networks with more sophisticated models of synaptic plasticity and delay as well as including several other small features, optimisations and bug fixes for certain system configurations. This release supports GCC >= 4.9.1 on Linux, Visual Studio >= 2013 on Windows and recent versions of Clang on Mac OS X.

1. Sparse synaptic connectivity can now be initialised using small snippets of code run either on GPU or CPU. This can save significant amounts of initialisation time for large models.
2. New 'ragged matrix' data structure for representing sparse synaptic connections -- supports initialisation using new sparse synaptic connecivity initialisation system and enables future optimisations.
3. Added support for pre and postsynaptic state variables for weight update models to allow more efficient implementatation of trace based STDP rules. See \ref sect34 for more details.
4. Added support for devices with Compute Capability 7.0 (Volta) to block-size optimizer.
5. Added support for a new class of 'current source' model which allows non-synaptic input to be efficiently injected into neurons.
6. Added support for heterogeneous dendritic delays.
7. Added support for (homogeneous) synaptic back propagation delays using SynapseGroup::setBackPropDelaySteps.
8. For long simulations, using single precision to represent simulation time does not work well. Added NNmodel::setTimePrecision to allow data type used to represent time to be set independently.

1. GENN_PREFERENCES::mergePostsynapticModels flag can be used to enable the merging together of postsynaptic models from a neuron population's incoming synapse populations - improves performance and saves memory.
2. On devices with compute capability > 3.5 GeNN now uses the read only cache to improve performance of postsynaptic learning kernel.

1. Fixed bug enabling support for CUDA 9.1 and 9.2 on Windows.
2. Fixed bug in SynDelay example where membrane voltage went to NaN.
3. Fixed bug in code generation of SCALAR_MIN and SCALAR_MAX values.
4. Fixed bug in substitution of trancendental functions with single-precision variants.
5. Fixed various issues involving using spike times with delayed synapse projections.