perf: eliminate temporary allocations in Philox rounds

[leejet]

Summary

• Replace temporary vectors in Philox rounds with 64-bit products and direct extraction of their high and low halves.
• Eliminate 327,680 small allocations per 16,384-value draw while preserving the random sequence and floating-point
  calculations.

Related Issue / Discussion

N/A

Additional Information

.\bin\Release\sd-cli.exe -m ..\..\stable-diffusion-webui\models\Stable-diffusion\v1-5-pruned-emaonly.safetensors -p "a lovely cat" -v --fa

12.45it/s -> 14.80it/s

Checklist

• I have read and confirmed this PR follows the contribution guidelines.