Reduce memory overhead from copy2d

[EndlessReform]

This PR fixes:

• Now using the official KVCache implementation
• For smaller models, the Candle repeat_interleave optimization with Tensor::cat is actually slower than the previous .unsqueeze().expand().reshape() even despite doing non-strided copies, due to the overhead of doing 8 sequential copy2ds + another to make it contiguous. I've reverted to the original, which improves speeds by another 20 tokens/sec. Further optimization is needed though, as this is down from 70µs to 25µs on a 4090 - still as much as the actual attention mechanism!
• Moved rep pen calculation for slow codebooks to GPU, to avoid copying from GPU -> CPU for rep pen -> GPU for softmax and temp scaling -> sampling. It's lower-precision at BF16, but shaves off another ~100µs since we have to do this 8 times per timestep.
• Fix RTF calculation for DualAR

For anyone else reading this: precomputed indexes aren't actually faster for interleaving the KV tensors, probably due to overhead.

Still behind the reference implementation with torch.compile: 160 tokens/sec on Nvidia 4090, vs 250 tokens/sec reported by maintainers. We're also getting increasingly CPU-bound due to kernel launch overhead.

Next steps:

• Add flash attention support
• Look into efficient RoPE and fused add+norm operators: they're making up a large percentage of runtime now
• Potential CustomOp for interleaving, it's still taking too long
• Experiment with CUDA graphs: looks like Candle may be adding support soon? Experimentations around cuda graphs huggingface/candle#2538