feat: support separate DiT/VAE service instances for Flux.

[z-jun03]

Description

Add role-based DiT/VAE separation for Flux inference services.

This change:

• Adds all, dit, and vae instance roles for Flux pipelines.
• Allows DiT instances to generate latent outputs without loading VAE weights.
• Allows VAE instances to load only VAE weights and serve latent decoding requests.
• Adds configurable VAE service endpoints through dit_vae_service_addresses.
• Adds DiT-to-VAE RPC communication through CommChannel and RemoteWorker.
• Supports round-robin VAE selection, health checks, request timeout configuration, and failover.
• Preserves the existing all mode for backward compatibility.
• Adds configurable DiT Worker service ports for multi-instance deployment.

Related Issues

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

Thank you for contributing to xLLM. Before requesting review, please make sure the following items are complete.

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Recommended title:

feat: support separate DiT and VAE service instances for Flux.

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: Flux DiT/VAE separated deployment and runtime validation have been completed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

• The implementation supports multiple DiT instances sharing one or more VAE service instances through configurable VAE endpoints.
• DiT instances send latent tensors to VAE instances through the existing Worker RPC interface.
• VAE instances return decoded image tensors through the RPC response.
• VAE selection uses local round-robin routing with health checks and failover when a VAE request fails.
• The existing dit_instance_role=all mode is preserved for backward compatibility.
• The current implementation uses static dit_vae_service_addresses; dynamic service discovery, a centralized router, and global load balancing are not included.
• DiT and VAE currently reuse the existing DiTWorkerImpl and WorkerService infrastructure, while model component loading and forward behavior are selected by dit_instance_role.
• Automated unit tests for the separated DiT/VAE deployment path are not included in this change; validation is currently based on NPU build and runtime testing.