This PR:

1. Drops PTX backend, and adds Metal on Mac self-hosted runner — removed all PTX matrix entries and PTX-only CUDA-graph steps from build-and-run.yml; added metal as a third backend that runs on [self-hosted, macOS] while opencl/cuda stay on [self-hosted, Linux], using a matrix.backend.os field to route runs-on per entry.
2. Provides a fix for Metal backend: avoid split-KV attention kernel Qwen3 crashes on Metal backend — root-caused and fixed the Qwen3 Metal crash: TornadoVM's Metal backend fails to JIT the multi-workgroup split-KV flash-decoding attention kernel (processHeadsFlashAttentionSplitKV) — confirmed by tracing crash logs down to kernel submitted: id=0x0, method=, argCount=0 right before the fatal "system integrity has been compromised" abort. Also tried the Qwen3-specific single-phase online-softmax kernel — same failure. The generic processHeadsFlashAttention kernel (already used by Llama, and known-working on Metal) does not have this problem. Fixed by making both Qwen3FP16FFNLayers and Qwen3Q8_0FFNLayers select that kernel on Metal only (mirroring the existing isWarpShuffleSupported() backend-detection pattern), skipping the now-unneeded combine phase.

Verified directly on hardware (not just built): Qwen3-0.6B in F16 and Q8_0, standard and --with-prefill-decode, plus a Llama regression check — all run cleanly on both 5.2.0-jdk21-metal and 5.2.0-jdk25-metal, producing real inference output at ~15-17 tok/s with no crash.