Very wide model support

[ptheywood]

Allow models with more than 128 wide layers to run without error, by not using a fixed number of CUDAStreamCompactionConfig.

Closes #727

Todo

• [x] Add tests which expose the current issue when a number of different features are used
• [ ] Malloc a dynamic array based on the width of the widest layer in the model
• [ ] CUDAScatter relies on MAX_STREAMS, so refactoring is required.
• [ ] FLAMEGPUDeviceException relies on MAX_STREAMS, so refactoring is required.
• [ ] Remove CUDAScanCompaction::MAX_STREAMS, replacing with a member variable of the current number allocated.
• [ ] Remove/adjust related exceptions (where CUDAScanCompaction::MAX_STREAMS is/was checked)
• [ ] Test with seatbelts
• [ ] Test without seatbelts

Notes

CUDAScanCompaction::MAX_STREAMS is hardcoded to 128, the upper limit that can run on a (<= SM75) device at once. This is a bad assumption.

Models can have more than 128 functions per layer, which requires that many streams

CUDAScatter is initialsed as a singleton member of CUDASimulation, so we know the fixed model properties at that point in time, so can add a call to allocate enough data then.

DeviceExceptionManager has an array of 1 device pointer to a DeviceExceptionBuffer per stream, and host memory to copy that back to. DeviceExceptionManager is a member of cudaSimulation::singletons, so can be allocated during singleton initialisaton.

CUDAScanCompaction is a member variable of CUDAScatter, which is default initialised (rather than being manually constructed or mentioned by an inisialiser list. This will need to be changed to pass the number of streams to create during conscruction, or to allocate the required number of elements later. This appeasr to to be the only instatntiations of CUDAScanCompaction afiak.

Destrcution / deleteion will also be required.