fix(build): filter InfiniOps adapters on every backend

[voltjia]

Summary

• apply INFINI_OPS_OPS adapter filtering on every backend instead of only inside the Cambricon + ATen setup block
• keep the Cambricon-specific torch_mlu ABI and include configuration scoped to Cambricon
• preserve the existing fallback that excludes the paged-attention InfiniOps adapter when no operator selection is provided

Root cause

The adapter-filtering block was nested in the Cambricon + ATen condition, so NVIDIA and other backends compiled every *_infiniops.cc adapter even when ops.json selected only a subset. A linked InfiniOps library could then fail to load because an unselected adapter retained an unresolved operator reference.

Validation

• git diff --check
• exercised this exact change in the stacked NVIDIA validation using accelerator-dev/nvidia:latest and an ops.json-selected InfiniOps build
• confirmed the unselected IndexSelect adapter no longer left an undefined InfiniOps reference
• built and installed infinicore_cpp_api and _infinicore; the resulting Python package imported successfully