ci: build release wheels against PyTorch 2.13

[b8zhong]

Builds the release wheels against PyTorch 2.13 (current latest) instead of the 2.4–2.8 matrix.

Changes

• publish.yml: torch-version matrix → ["2.13.0"].
• publish.yml: drop Python 3.8/3.9 — torch 2.13 is requires-python >= 3.10. The torch 2.4.0 x py3.13 exclude is dropped with it, since 2.4.0 is no longer in the matrix.
• _build.yml: add a 2.13 entry to the min/max CUDA lookup (126 / 129). Without it the step raises KeyError: '2.13'. With cuda-version: 12.9.1 the build resolves to the cu129 wheels, which exist for 2.13.0 (available variants: cu126 / cu129 / cu130 / cu132).

Notes

• Not built or otherwise verified — the wheel job only fires on v* tag creation, so nothing runs from this PR.
• Pre-existing and untouched: the build step pins setuptools==75.8.0, while torch 2.13 declares setuptools>=77.0.3. pip will flag the conflict. Happy to bump the pin here if you'd prefer.

[b8zhong]

Wrong target — sglang consumes the sgl-deep-gemm wheel built by sglang's own release-whl-deepgemm.yml (TORCH_VER in that workflow + docker/sgl-deep-gemm.Dockerfile), not this repo's publish.yml. Redoing the bump there instead.