ci(container): publish SenseVoice runtime to GHCR

[LauraGPT]

Summary

• publish a model-weight-free linux/amd64 runtime image to GHCR from main, version tags, and manual runs
• build without pushing on pull requests and provide immutable commit tags
• align the pinned PyTorch/CUDA base with the repository torch floor
• document the official image, persistent model cache, real port 50000, and CPU/GPU launch paths in three languages

Verification

• red/green container contract: 3 expected failures on main, then 3/3 passed
• Python unittest and pytest runners both pass
• git diff --check passes
• exact-head GitHub Actions must complete the full container build before merge

Related: #339

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com