Don't keep box muller transform state between kernel launches

[neworderofjamie]

This builds on the new hybrid HIP/CUDA backend from #647 so review that first!

As described in #648, beyond its (arguably excessive) 192 bit of state, the curand XORWOW RNG used to provide randomness for neuron and custom connectivity updates also stores 160 bits of box muller transform state in the curandState struct (BM draws two numbers and produces two normally distributed values so this state is used to cache one of those results for subsequent calls to curand_normal).

In this PR, when using CUDA and HIP backends, we create our own XORWowStateInternal struct in definitions.h without the BM state and store this in memory. At the start of the neuron and custom connectivity update kernels, we copy the fields from the XORWowStateInternal struct into a local curandState and, at the end, we copy them back.

Excitingly, because we are very memory bandwidth bound, this makes the neuron kernel on the cortical microcircuit model about 60% faster. On my A5000 (running for 1 second):

	Before[seconds]	After [seconds]
neuron update	0.18	0.11
presynaptic update	0.23	0.24

[codecov]

Codecov Report

Attention: Patch coverage is 98.66667% with 1 line in your changes missing coverage. Please review.

Project coverage is 88.97%. Comparing base (54101ca) to head (beb19d4).

Files with missing lines	Patch %	Lines
include/genn/genn/code_generator/environment.h	96.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##              hip     #649      +/-   ##
==========================================
+ Coverage   88.91%   88.97%   +0.05%     
==========================================
  Files         108      108              
  Lines       14717    14768      +51     
==========================================
+ Hits        13086    13140      +54     
+ Misses       1631     1628       -3     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.