fix(ci): update lmdeploy commit in Dockerfile and fix some ci image version problems

[CyCle1024]

This pull request updates the Docker build and deployment workflow to use newer dependencies and improves the way Docker images are tagged and pushed. The main changes involve upgrading the default PyTorch version, adjusting LMDeploy installation, updating image registry handling, and refining image tagging conventions.

Dependency and version upgrades:

• Updated the default PyTorch version to 2.9.1 in the workflow, image build script, and Dockerfile, replacing the previous 2.8.0/2.9.0 defaults.
• Changed LMDeploy installation in the Dockerfile to use the official pip package for a specified version or build from a custom URL, and switched from using LMDEPLOY_VERSION to LMDEPLOY_URL as the main build argument.

Docker image registry and tagging improvements:

• Removed the hardcoded registry prefix from IMAGE_NAME and now prepend the registry from secrets during the push step, improving flexibility and security.
• Refined tagging: introduced SHORT_DOCKER_TAG for clarity and updated SCHEDULE_TAG to reflect the new PyTorch version.

Configuration and image usage updates:

• Updated the default Docker image used in autotest/config.yaml to ailab-llmrazor/xtuner:pt29_latest, ensuring tests use the latest image built with the new workflow.

Dependency management:

• Added the tilelang package to the Dockerfile's installation list, ensuring it is available in the build environment.