Switch demo to viser server mode — render frames GPU-side, no client WebGL

Install prebuilt CUDA wheels — no compilation on the HF builder (fixes OOM)

Patch viser version check — HF's HTTP/2 proxy drops the WS subprotocol

Drop +PTX from CUDA arch — shrink fused_knn_attn compile under HF builder RAM

Use Python 3.12 (deadsnakes) — optgs needs PEP 695 generic syntax

Fix build OOM: compile CUDA for A10G arch only, cap MAX_JOBS

Add Docker-based Learn2Splat demo (viser GUI)