Add a CIFAR-10 demo with SET/RigL comparison

[DarshanFofadiya]

Context

All current DST algorithm comparisons (SET vs RigL vs Static) run on MNIST. CIFAR-10 is the next standard benchmark and is where the RigL paper shows its advantage over SET most clearly.

What needs to happen

1. Write examples/demo_17_cifar10.py following the pattern of demo_11_rigl_vs_set_vs_static.py.
2. Use a simple CNN or ResNet-20 (not an MLP — CIFAR needs convolutions to be meaningful). The SparseLinear layers would replace the fully-connected layers; conv layers stay dense for v0.1.
3. Compare Static vs SET(0.1) vs RigL(0.1) at 90% sparsity on the FC layers.
4. Report accuracy and training curves.

Notes

• CIFAR-10 downloads via torchvision.datasets.CIFAR10 — same pattern as MNIST in demo_05.
• This is a good first issue because the infrastructure (SparsityAlgorithm, SparseLinear, training loop pattern) all exists. You're writing a new demo script, not new library code.

Files to read

• examples/demo_11_rigl_vs_set_vs_static.py — the MNIST version to use as a template
• examples/demo_05_mnist.py — data loading pattern