feat!: [codex] support polymorphic handles and configs

[voltjia]

Summary

• Add a reusable Cloneable<Base, Derived> CRTP helper for concrete Config and Handle subclasses.
• Give Config and Handle virtual destructors and virtual Clone() methods while keeping both base classes concrete.
• Replace the two sliced OperatorBase value members with owned std::unique_ptr clones; existing set_config(const Config&) and set_handle(const Handle&) signatures stay unchanged.
• Add a focused installed-header C++ smoke test covering derived state, base state, clone independence, and source-object lifetime.

Motivation

Operator-specific options currently have to be added to the shared src/config.h, because assigning a derived Config or Handle to OperatorBase slices its dynamic state. This PR provides only the ownership mechanism needed for operator-specific subclasses.

This change is deliberately independent of #800 and contains no Triton/JIT, cache-key, constructor-dispatch, or handle-borrowing optimization changes.

Type of Change

• feat - polymorphic Config and Handle ownership
• fix - bug fix
• perf - performance improvement
• refactor - code restructuring without behavior change
• test - focused C++ API coverage
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

The shared public types and OperatorBase are used by every backend.

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

Smoke Test Result

Validated from commit 8142a511cf9cdac79d31f2247750a3344c2f7c20:

g++ -std=c++17 -Werror -Isrc -fsyntax-only <generic Cloneable probe>
# passed

python3 -m py_compile tests/test_cpp_api.py
# passed

clang-format 21.1.8 --dry-run --Werror \
  src/cloneable.h src/config.h src/handle.h src/operator.h
# passed

ruff 0.15.22 format --check tests/test_cpp_api.py
ruff 0.15.22 check tests/test_cpp_api.py
# passed

GitHub Actions provides the installed-library and backend build/run coverage. All required checks passed, including clang-format, Ruff, documentation build, and unit plus shadow jobs for NVIDIA, Iluvatar, MetaX, Moore, Cambricon, and Ascend.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	yes	local C++17 API syntax probe passed	installed-library smoke covered by backend CI
NVIDIA	yes	unit and shadow CI passed	passed
Iluvatar	yes	unit and shadow CI passed	passed
MetaX	yes	unit and shadow CI passed	passed
Cambricon	yes	unit and shadow CI passed	passed
Moore	yes	unit and shadow CI passed	passed
Ascend	yes	unit and shadow CI passed	passed

Benchmark / Performance Impact

No performance optimization is included.

Config is cloned when an operator is constructed, which normally happens on a cache miss. The existing invocation path calls set_handle() for every invocation; after this change that performs one virtual Clone(), one derived-object copy, and one heap allocation per call. A borrowed-handle fast path can be evaluated separately without coupling it to the polymorphism API.

Notes for Reviewers

• This is ABI-breaking: Config and Handle gain virtual tables, and the protected OperatorBase::config_ / handle_ members change from values to std::unique_ptr.
• Concrete CRTP subclasses must be final; Cloneable enforces this when Clone() is instantiated.
• The base Config and Handle classes remain directly constructible and clone themselves by default.
• There is no production RTTI, dynamic type hash, or config fingerprint in this PR.
• Cache-key behavior is unchanged. An operator whose derived Config fields affect cached behavior must include those fields explicitly in its existing CacheKeyBuilder<Key> specialization.
• set_config() and set_handle() remain replaceable, matching the previous setter semantics; no one-time initialization rule is added.