fix: honor virtualenv exemption in engine matching for multimodal LLM, embedding and rerank

[OliverBryant]

Problem

On an environment without vLLM/SGLang installed (or without a local GPU), a
chat-only model like qwen3 can be launched with vLLM/SGLang, but their
multimodal counterparts cannot — the engines don't appear as selectable in the
WebUI. The same asymmetry affects Qwen3-VL-Embedding and Qwen3-VL-Reranker.

Engine availability is computed by get_engine_params_by_name_with_virtual_env,
which calls each engine's match_json. Some engines put runtime-only checks
— library installed, specific version, CUDA GPU, Linux — inside match_json,
before (or instead of) the virtualenv exemption. Chat LLM engines already had
the exemption; the multimodal / embedding / rerank vLLM paths did not, so a
declared engine was hidden at listing time even though virtualenv installs it on
demand at launch (check_engine_by_spec_parameters_with_virtual_env already
bypasses this).

Fix

Make the virtualenv exemption consistent: when virtualenv is enabled,
environment-only checks (library present, version, GPU, OS) are skipped so an
architecture/format-compatible engine is still offered and installed on demand.
When virtualenv is disabled, all strict checks are preserved.

LLM

• vLLM: exempt GPU/OS/library checks in VLLMModel and VLLMMultiModel.
• SGLang: add _virtual_env_allows_missing_sglang; exempt GPU/OS/library checks
  in the generate, chat and vision branches.
• LMDeploy: add _virtual_env_allows_missing_lmdeploy; exempt the library check
  in the chat branch.

Embedding / Rerank

• Add a shared virtual_env_allows_missing_engine() helper in model/utils.py.
• embedding vLLM: exempt the "vLLM is not installed" and version checks for
  Qwen3-VL-Embedding under virtualenv.
• rerank vLLM: replace the unconditional requires vLLM>=0.14.0 rejection for
  Qwen3-VL-Reranker with a real import + version check that is exempt under
  virtualenv, so the engine is offered instead of being permanently disabled.

Unchanged (already correct): MLX (Apple-silicon-only, marker auto-dropped on
non-macOS), Transformers / llama.cpp, embedding/rerank
sentence_transformers/flag/llama.cpp, and image (stable_diffusion) — all keep
library checks in check_lib, so listing already honors the exemption. Audio /
video do not use the engine-selection mechanism.

Tests

• LLM: regression test asserting qwen3.5 offers vLLM/SGLang/Transformers/
  llama.cpp under virtualenv while MLX/LMDeploy (not declared) are not listed.
• Embedding: strengthened to assert Qwen3-VL-Embedding-2B offers vLLM.
• Rerank: new engine-params test asserting Qwen3-VL-Reranker-2B offers vLLM.

All new/updated assertions fail on the previous code and pass with this change.
Verified manually for a no-GPU host and a simulated CUDA+Linux host with the
libraries absent.

[gemini-code-assist]

Code Review

This pull request updates the LMDeploy, SGLang, and vLLM engines to tolerate missing libraries, GPUs, or OS requirements when virtualenv mode is enabled, allowing these engines to be listed and installed on demand. It also adds a regression test for multimodal engines under virtualenv. The review feedback suggests simplifying the implementation by importing the XINFERENCE_ENABLE_VIRTUAL_ENV constant directly at the top level of the modules instead of defining redundant helper functions with try-except blocks.

[OliverBryant]

Thanks for the review. Addressed the feedback: removed the redundant _virtual_env_allows_missing_* helper functions in sglang/core.py and lmdeploy/core.py and now import XINFERENCE_ENABLE_VIRTUAL_ENV at the module top level, using it directly (confirmed no circular-import risk since constants only depends on os/pathlib). The shared virtual_env_allows_missing_engine() helper in model/utils.py (used by the embedding/rerank vLLM paths) was likewise simplified to reference the top-level constant. Left the pre-existing _virtual_env_allows_missing_vllm in vllm/core.py unchanged to keep the diff focused. Also rebased onto the latest main to drop unrelated commits.

[qinxuye]

Two correctness issues remain before this can be approved.

[qinxuye]

Virtualenv can install Python packages, but it cannot add accelerator hardware or turn a non-Linux worker into Linux. Since XINFERENCE_ENABLE_VIRTUAL_ENV defaults to true, this makes vLLM/SGLang match on unsupported workers; on a macOS host with no CUDA, VLLMMultiModel.match_json() now returns True. It also ignores an explicit request-level enable_virtual_env=False because match_json reads the process-global constant. Please keep the hardware/OS checks unconditional and use the effective request flag only to exempt missing runtime libraries. The same issue applies to the SGLang branches changed in this PR.

[qinxuye]

This assumes the virtualenv will install a compatible vLLM, but these Qwen3-VL model specs only declare #vllm_dependencies#, which currently expands to vllm>=0.11.2. An existing environment with 0.11.2 already satisfies that requirement, while model initialization later requires vLLM>=0.14.0 and fails. Please encode vllm>=0.14.0 in the Qwen3-VL embedding and reranker virtualenv requirements (and test the expanded requirement) before bypassing this version check.