docs: clarify model-specific network batching

[anHappyDog]

Summary

Correct the English and Chinese READMEs' blanket batch-size-1 description and keep the documentation homepages in sync. π0.5 supports cross-session batching through HTTP and WirelessComm, with a default batch limit of 1; DM0.5 and StreamVLN require --max-batch 1.

Explain the shared-client example's batch limit and collection window: partial batches execute when the window expires, and the window excludes queueing and inference time. Preserve ActiveVLN's experimental status and pending real-checkpoint GPU parity notice. Use GitHub's rendered anchor for the Chinese README's shared-client guide.

Link robot deployment to the existing English and Chinese EmbodiRun support matrices on main, and align the Chinese documentation navigation with the English README.

Verification

Checks ran in isolated Python environments:

• mkdocs build --strict and mkdocs build --strict -f mkdocs.zh.yml — both passed.
• python -m pytest tests/test_serving_batches.py -q -m 'not gpu' — 11 passed, 1 GPU test deselected.
• ruff check embodiinfer tests and ruff format --check embodiinfer tests — both passed.
• git diff --check — passed.
• All 120 local and same-repository links in the changed documents resolve to existing paths and declared or generated anchors; the Chinese shared-client anchor was also verified against GitHub's rendered HTML.
• Both HTTP and WirelessComm server --help commands expose the documented batch flags.
• Both EmbodiRun support-matrix targets exist on main; this PR can merge independently of EmbodiRun's companion documentation PR.

Documentation only; no model execution changes or new GPU/real-checkpoint validation claims.