[Fix] Align local-eval deps with ext-prod; default skip-reference

[yuanzhang-us]

Make local docker eval match the dependency stack used by the production sol_execbench_external image so externally-authored solutions compile and run identically.

• pyproject.toml: pin torch==2.9.0 (was >=2.10.0); drop unused torchvision dep that blocked the torch downgrade. nvidia-cutlass-dsl stays at 4.4.1.
• uv.lock: regenerated; torch 2.10->2.9, triton 3.6->3.5, nvidia-* libs aligned with torch 2.9.
• docker/Dockerfile: add UV_HTTP_TIMEOUT=600 so cold builds don't fail on the large nvidia-cutlass-dsl-libs-cu13 wheel under the default 30s cap.
• cli/main.py: rename PYTORCH_ALLOC_CONF -> PYTORCH_CUDA_ALLOC_CONF in the eval subprocess env. torch 2.9 only honors the CUDA form, so the intended expandable_segments allocator never engaged before this fix.
• core/bench/config/benchmark_config.py: flip benchmark_reference default from True to False. Reference implementations can take >1h for some problems and dominate evaluation time; users that need a speedup factor can re-enable via --config.
• bench_config.example.json: ship a template containing every field at its default so users have a copy-and-edit starting point.
• README.md: document the BenchmarkConfig fields and the new example template.