[diffusion] cli: support component attention backend overrides

[mickqian]

Summary

• Add --component-attention-backends for per-component native attention backend overrides.
• Apply overrides while pipeline-defined modules/components are constructed, without hardcoding model-specific components.
• Carry component name through attention backend selection logs.
• Document the new CLI/API usage in both docs trees.

Test

• git diff --check
• Local unittest collection is blocked on this Mac by ImportError: cannot import name '_cuda_beginAllocateCurrentThreadToPool' from 'torch.cuda.memory'.

[gemini-code-assist]

Code Review

This pull request introduces the --component-attention-backends CLI argument, allowing users to specify per-component attention backend overrides (e.g., text_encoder=torch_sdpa). The implementation includes updates to the attention selector to support ContextVar-based overrides, integration into component loaders and pipeline construction, and logic in ServerArgs to parse both comma-separated strings and dotted CLI arguments. Documentation and unit tests have been updated accordingly. I have no feedback to provide.

[mickqian]

/tag-and-rerun-ci