refactor: finish removing vvla-prefixed identifiers

[hootandy321]

Summary

Finishes the vvla to embodiinfer rename. The package, imports, distribution, and
documentation were already renamed; this removes the last vvla-prefixed surfaces so
the repository has a single consistent name.

• Wire schemas: vvla.policy.*.v1, vvla.error.v1, and vvla.activevln.reference become
  embodiinfer.policy.*.v1, embodiinfer.error.v1, and embodiinfer.activevln.reference.
• Console scripts: the vvla-serve / vvla-http-serve / vvla-wireless-serve aliases are
  removed; only embodiinfer-serve, embodiinfer-http-serve, embodiinfer-wireless-serve remain.
• Test gates: VVLA_* environment variables (checkpoint/reference paths, compile-cache
  bootstrap assertion) become EMBODIINFER_*.
• On-disk layout: vvla-execution-entries, vvla-content-blobs, vvla-quarantine,
  .vvla-compile-cache.lock, and ~/.cache/vvla/pi05 become their embodiinfer names.
• Runtime fused-weight cache attributes _vvla_fused_* become _embodiinfer_fused_*
  (process-local module.__dict__ entries, never serialized in state dicts).
• The vvla import shim, the deprecated Vvla/VvlaError aliases, and the
  vvla_eager/vvla_sdpa benchmark modes are removed; the R2R manifest schema string
  and prepared-bundle naming follow the new prefix.
• Docs, README, benchmarks, and historical report paths use the new names.

Breaking: clients that send vvla.policy.* schemas or invoke vvla-*-serve must move to
the embodiinfer.policy.* schemas and embodiinfer-*-serve commands.

Verification: full CPU suite (437 passed, 1 pre-existing transformers-dependent failure,
52 skipped), ruff check and ruff format --check on embodiinfer, tests,
benchmarks, scripts, uv sync --frozen, and git grep -i vvla returns nothing.

[anHappyDog]

Reviewed commit: 1d4221ab3d632be73cbb98df4713e642067e00a4.

[P2] Complete the RLinf integration migration before removing the compatibility shim

Locations: pyproject.toml:142–143 and docs/en/serving.md:188–191. The Simplified Chinese serving guide has the same issue.

This PR removes the vvla import shim, but adapter.py on the public RLinf repository's vvla-rollout-backend branch still executes from vvla import make_policy. After installing this PR in a temporary environment, calling the existing adapter's make_vvla_policy('vvla') reproduces ModuleNotFoundError: No module named 'vvla'.

The updated serving guide also points to an embodiinfer-rollout-backend branch in BUAA-CI-LAB/RLinf, but that branch does not currently exist and the link returns 404. The newly documented embodiinfer_gr00t model type is also rejected by the existing adapter; the integration still registers names such as vvla_gr00t. Renaming references in this repository does not update the downstream branch or model registrations.

Please migrate and publish the corresponding RLinf imports and model registrations before removing the compatibility shim, then update both serving guides to reference an available branch and valid configuration names.

Validation: full CPU suite 476 passed, 61 skipped (Python 3.13, Torch 2.13.0+cpu, Transformers 4.57.1). Ruff lint, format, and diff checks passed. GPU and real checkpoint validation was not run.