Reduce register pressure in one-workgroup radix sort

[danhoeflinger]

Replace cached uint32_t* pointer array (__counters[__block_size]) with a uint16_t bin index array (__bins[__block_size]) in the one-workgroup radix sort kernel. This halves the register footprint of the cached data (32 vs 64 registers on 64-bit pointer platforms) by storing only the bin index and reconstructing the counter address in the post-scan loop.

The pointer array existed to avoid recomputing each element's bucket after the scan phase, but caching full 64-bit pointers is expensive in register-constrained kernels. Storing the bin index preserves the benefit (no recomputation from values) at half the register cost.

I've also taken the opportunity to add std:: to uint*_t within this function.

Note: This is within margin of error from a performance perspective, but for some driver versions, this avoids generating kernels with register spill warnings in real use cases.

[Copilot]

Pull request overview

This PR reduces register pressure in the one-workgroup radix sort SYCL kernel by replacing a per-element cached counter-pointer array with a per-element cached bin-index array, reconstructing the counter address only when needed after the scan.

Changes:

• Replace uint32_t* __counters[__block_size] caching with std::uint16_t __bins[__block_size] caching and recompute the counter address in the post-scan step.
• Apply std:: qualification to uint*_t types within this kernel function for consistency.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[timmiesmith]

LGTM. Thank you.