GPU accelerated backend for interpolation functions

[AndrewHarris709]

Includes a GPU accelerated backend for all the currently available interpolation functions. The GPU implementation uses CUDA kernels created with numba. Using this backend results in a noticeable performance boost in large datasets.

A CUDA compatible graphics card (with proper drivers) is required for the GPU backend to function properly.

The backend used by interpolation functions can be changed in two ways:

1. The backend argument can be provided to any render/interpolation function, set this to gpu for the CUDA backend.
2. The backend property of a SarracenDataFrame can be set. Any interpolation/render function used with this dataset will use the defined backend.

An example of performance differences between backends: 2D Rendering - Small dataset (2.62MB)

• cpu: 274ms
• gpu: 748ms

2D Rendering - Large dataset (1.31GB)

• cpu: 7.14s
• gpu: 1.78s

Testbench:

• CPU: AMD Ryzen 9 3900X
• GPU: NVIDIA GeForce RTX 2080 Super

As well, the CPU backend has been modified to remove a critical race condition. CPU interpolation should now complete without noise.

[AndrewHarris709]

Atomic operations have been reverted. With this change the unit test test_race_conditions() will fail, this is expected.

[ttricco]

@AndrewHarris709 Are you able to take a look at resolving the merge conflicts here?