fix(converter): detect SGLang device backend

[Le8r0nJames]

What does this PR do?

This PR makes the SGLang weight converter use AWEX's shared runtime device
detection instead of inferring the device backend from visibility environment
variables.

Explicit device_backend/device_type configuration and the HCCL backend
continue to take precedence. When they are absent, the converter now delegates
to device_util.get_device_type().

Some launchers export both CUDA_VISIBLE_DEVICES and
ASCEND_RT_VISIBLE_DEVICES. Previously, the presence of the Ascend variable
caused a CUDA process to be misclassified as NPU, which could incorrectly apply
NPU-specific weight conversion behavior.

A regression test covers the case where both visibility variables are present
while CUDA is the available runtime device.

Related issues

None.

Does this PR introduce any user-facing change?

No public API or binary protocol changes. Device backend selection now reflects
the configured or actually available runtime device instead of relying on
visibility environment variables.

• Does this PR introduce any public API change?
• Does this PR introduce any binary protocol compatibility change?

Validation

• ruff check awex/converter/sglang_converter.py awex/tests/test_qwen3_moe_sglang_converter.py
• ruff format --check awex/converter/sglang_converter.py awex/tests/test_qwen3_moe_sglang_converter.py
• pytest awex/tests/test_qwen3_moe_sglang_converter.py awex/tests/test_qwen3_dense_sglang_converter.py awex/tests/test_meta_resolver.py awex/tests/test_model_registry.py -q
  • 30 passed
  • 2 skipped