ci(container): настроить сборку образа

[Fgeeha]

Closes #379

Summary

• add a CUDA development-based Containerfile for the ft serve image
• add a GitHub Actions workflow that builds the image for pull requests and manual runs
• publish GHCR images only for pushes to FlashML-org/FreeToken
• publish latest from the default branch plus semver and commit-SHA tags
• document how to run the image with a local model directory

Publishing behavior

Pull requests and manual workflow runs build without publishing.

Pushes to main and version tags publish to ghcr.io/flashml-org/freetoken.

The publish job uses GITHUB_TOKEN with packages: write and is restricted to the upstream repository.

Validation

Test environment:

• OS: local Linux development environment
• GPU / driver / checkpoint: not applicable; this change validates container build configuration and does not execute
  model inference

Commands run:

  go run github.com/rhysd/actionlint/cmd/actionlint@v1.7.7 .github/workflows/container-image.yml
  docker build --check -f Containerfile .
  git diff --check

All commands completed successfully.

A full container build was started but not completed because it requires downloading the multi-gigabyte CUDA base
image. No successful full image build or inference test is claimed.