SparkInfer restores Windows downloads on SN74

This is a distribution win more than a performance breakthrough: optimized inference code only matters when developers can reliably install it.

• –v0.6.34 fixes missing Windows fallback functions that caused linker failures during packaging. [Tao Outsider](https://www.taooutsider.com/blog/gittensor-sn74-sparkinfer-windows-release/)
• –Windows users gain a supported package, but without the NVFP4 acceleration path available in other builds.
• –SparkInfer targets Blackwell GPUs and reports strong throughput against llama.cpp on RTX 5090 workloads, though those benchmarks remain hardware- and model-specific. [SparkInfer GitHub](https://github.com/gittensor-ai-lab/sparkinfer)
• –The release exposes why cross-platform packaging, CI coverage, and checksums are as important as kernel-level optimization.