fix: preserve implementation lookups under optimization

[voltjia]

Summary

• This PR emits explicit specializations for generated active implementation queries and default implementation lookups.
• This PR keeps the generic header-only lookup path available for operators that are not part of the generated installed API.
• This PR compiles the installed Add C++ smoke consumer with optimization enabled to cover both query and configless call paths.

Motivation

The explicit instantiations added in #924 do not prevent a compiler from inlining the class-body definitions in an optimized downstream translation unit. Because downstream code does not include backend implementation headers, an optimized consumer can incorrectly observe an empty implementation list and select no default implementation even though the installed library contains the implementations.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• 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 (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• 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

CPU Release library build passed.
Installed CPU C++ consumer compiled with -O3 and executed Add query, explicit-config Call, and configless Call successfully.
Installed CPU -O3 query consumer reported Add: 0, Gemm: 0, Mul: 0, Relu: 0, RmsNorm: 0, and Silu: 0.
NVIDIA Release library build passed with the generated specialization source compiled by NVCC.
Installed NVIDIA -O3 query consumer reported Add: 0, Gemm: 0 1, Mul: 0, Relu: 0, RmsNorm: 0, and Silu: 0.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	Passed	The 26 wrapper generator tests and the focused installed C++ API regression passed.
NVIDIA	Yes	Passed	The NVIDIA library and installed optimized query consumer passed on an NVIDIA A100.
Iluvatar	Yes	Not run	Hardware is unavailable in the current environment. Standard CI validation is required.
MetaX	Yes	Not run	Hardware is unavailable in the current environment. Standard CI validation is required.
Cambricon	Yes	Not run	Hardware is unavailable in the current environment. Standard CI validation is required.
Moore	Yes	Not run	Hardware is unavailable in the current environment. Standard CI validation is required.
Ascend	Yes	Not run	Hardware is unavailable in the current environment. Standard CI validation is required.

Full `pytest` output.

Running 26 items in this shard
..........................                                               [100%]
26 passed in 0.26s

Running 1 items in this shard
.                                                                        [100%]
1 passed, 6 deselected in 1.20s

Benchmark / Performance Impact

N/A. This PR changes where generated implementation lookup code is instantiated and does not change operator execution paths.

Notes for Reviewers

Please focus on the explicit specialization declarations in the installed generated header and their definitions in the generated library source. The generic helper templates intentionally remain in operator.h so non-generated custom operators preserve the existing header-only behavior.