fix(ops): select active implementation for configless calls

[voltjia]

Summary

• Make configless C++ Operator::Make and Operator::Call select the first active implementation for the input device.
• Resolve that default index inside the built library, where backend specializations are visible to installed-header consumers.
• Cover tensor, tensor-list, return-value, slot 0, and nonzero-slot paths without changing explicit Config behavior.

Motivation

Generated Python bindings already choose the first active implementation when a caller omits implementation_index. The configless C++ convenience overloads instead constructed Config{}, which always selected slot 0. Operators implemented only by a linked provider at slot 16 therefore failed even though a valid implementation was active.

The default cannot be derived from ActiveImplementations in a downstream translation unit because installed public headers do not expose backend specializations. This change emits one scalar resolver instantiation per operator in an implementation-visible generated source file and only declares it in the installed header.

N/A - no linked issue.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• Hygon (WITH_HYGON)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Remote environment: ssh nvidia, accelerator-dev/nvidia:latest, PyTorch 2.10.0a0+b4e4ee8.

INFINI_RT_ROOT=/opt/infinirt scripts/dev/build.sh nvidia --smoke --jobs 16
[build] done

python3 -m pytest tests -m smoke -q
123 passed, 18 skipped, 13904 deselected

The smoke allowlist is unchanged.

Focused Test Results

tests/test_generate_wrappers.py
25 passed

Installed <infini/ops.h> configless C++ consumer test
1 passed

The installed consumer test uses the real generated Abs slot 8 path for configless Make and Call. Existing Add coverage exercises slot 0 and CallReturning; a test-only slot 16 operator covers tensor-list descriptor construction.

The remote consumer compiler temporarily suppressed deprecation warnings emitted by legacy operators already present in the smoke-generated umbrella header. No deprecation behavior is changed in this PR.

Benchmark / Performance Impact

The default implementation query returns a scalar and runs only while entering a configless API. It does not allocate a vector and does not change kernel execution.

Notes for Reviewers

• Only overloads without a Config derive a default from the device's active implementations.
• Make(config, ...) and Call(handle, config, ...) still honor the exact requested index and never fall back.
• Configless Make keeps its public forwarding-reference signature and normalizes only the internal generated-factory call to the existing const T& explicit instantiations.
• CallReturning is covered because it delegates to configless Call after allocating the return tensor.