refactor(nvidia)!: link vLLM awq marlin repack

[voltjia]

Summary

• Replace the copied NVIDIA AWQ Marlin repack kernel with the linked vLLM _C::awq_marlin_repack provider.
• Preserve the public InfiniOps signature and output-tail convention through a thin Torch/C10 adapter.
• Remove 370 lines of duplicated CUDA implementation from the repository.

Motivation

awq_marlin_repack is already available in the installed vLLM provider. Calling
its registered C10 operator avoids maintaining a second copy of the upstream
kernel and keeps provider architecture support with the provider build.

This PR depends on #910 so configless C++ calls select linked slot 16 when it is
the first active implementation.

N/A - no linked issue.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement
• refactor - code restructuring
• test - existing operator tests validate the replacement
• docs - documentation only
• build / ci - build capability changes
• 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: A100 with a compatible installed vLLM provider and an
operator-pruned Release build using WITH_LINKED=ON.

python3 -m pytest tests -m smoke --devices nvidia -q
67 passed, 20 skipped, 6076 deselected

The smoke allowlist is unchanged.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Release operator-pruned linked build and existing smoke passed on A100	Focused AWQ provider tests: 13 passed
Iluvatar	No	N/A	N/A
MetaX	No	N/A	N/A
Cambricon	No	N/A	N/A
Moore	No	N/A	N/A
Ascend	No	N/A	N/A

Benchmark / Performance Impact

N/A - provider replacement; no performance claim.

Notes for Reviewers

• The adapter uses the caller-provided stream through the existing C10 stream bridge.
• The provider return tensor is copied into the existing output-tail tensor, so the public InfiniOps signature is unchanged.
• The resolver validates the exact operator schema and CUDA dispatch key before compilation.
• This intentionally makes NVIDIA AWQ provider-only: builds with
  WITH_LINKED=OFF no longer compile an AWQ implementation. Users must enable
  linked operators and install a compatible vLLM distribution.

BREAKING CHANGE: NVIDIA awq_marlin_repack now requires WITH_LINKED=ON and an installed compatible vLLM provider.