More tasks (FFT, matrix multiply, sorting networks)

[SolidRegardless]

Summary

Expand the benchmark task suite with three new GPU kernel implementations: Fast Fourier Transform (FFT), matrix multiplication, and sorting networks.

Motivation

The current five tasks demonstrate a range from bandwidth-bound to compute-bound workloads, but they're all element-wise operations with no inter-element dependencies. FFT, matrix multiply, and sorting networks introduce fundamentally different parallel patterns — butterfly networks, tiled shared-memory access, and compare-and-swap — that stress different aspects of GPU architecture and reveal new performance characteristics.

Acceptance Criteria

• Implement FftTask — a radix-2 Cooley-Tukey FFT on a power-of-two input array
• Implement MatrixMultiplyTask — dense matrix multiplication (e.g. 1024×1024 × 1024×1024)
• Implement SortingNetworkTask — a bitonic sort or odd-even merge sort on the GPU
• All three tasks implement IBenchmarkTask and integrate with the existing harness
• Each task includes proper host↔device memory transfers (no shortcuts)
• Results should be validated for correctness against a CPU reference implementation
• Performance should be benchmarked and reported alongside existing tasks

Technical Notes

• FFT requires shared memory and butterfly-pattern access — may need ILGPU shared memory primitives
• Matrix multiply benefits hugely from tiled/blocked algorithms using shared memory
• Sorting networks are interesting because they have a fixed comparison pattern, making them ideal for GPU parallelism despite sorting being traditionally sequential
• Batch size semantics differ per task: for FFT it's the transform length, for matrix multiply it's the matrix dimension, for sorting it's the array length
• Consider adding these under src/GpuStreamingDemo.Core/Tasks/ following existing patterns