ExLlamaV3 adds DFlash quantization, kernels

This is the kind of release that compounds. No single feature is flashy, but the combination of quantization support, kernel work, and stall fixes is exactly how inference stacks win on real workloads.

• –DFlash now moves from draft-model optimization into the quantization pipeline, which should make the speed path more practical for wider model deployment
• –Reducing autotune stalls matters because local inference libraries often burn time in setup and kernel selection, not just raw compute
• –Gemma 4-specific Triton kernels show ExLlamaV3 is still chasing architecture-level wins instead of relying on generic CUDA shortcuts
• –The release cadence from May 2 to May 9 signals an aggressively active maintainer loop, which is a real advantage in fast-moving open-source infra
• –The strongest story remains coding and agentic workloads, where earlier DFlash benchmarks showed the biggest throughput gains