feat(linked): support TVM FFI providers

[voltjia]

Summary

• Add a tvm_ffi linked-operator transport with .cu adapters, PEP 440 package-version checks, installed include discovery, and explicit DSO dependencies.
• Partition Torch and TVM-FFI adapter sources in the generated CMake manifest, and retain declared dependency DSOs with a scoped --no-as-needed.
• Document the transport contract and cover version, source selection, include discovery, dependency validation, and manifest generation.

Motivation

This is the transport-only prerequisite for reducing #930 to a FlashInfer provider PR. It intentionally contains no FlashInfer provider implementation, sampling operator changes, or third-party binaries.

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

The real integration check overlaid the follow-up FlashInfer provider files only in an isolated NVIDIA validation tree. Those files are not present in this PR.

python3 -m pytest tests/test_resolve_linked_ops.py -q
Running 37 items in this shard
..................................... [100%]
37 passed in 0.94s

python3 -m ruff check scripts/resolve_linked_ops.py tests/test_resolve_linked_ops.py
All checks passed!

cmake -S /src -B /work/cmake-build-linked   -DCMAKE_BUILD_TYPE=Release   -DINFINI_RT_ROOT=/opt/infinirt   -DAUTO_DETECT_DEVICES=OFF   -DAUTO_DETECT_BACKENDS=OFF   -DWITH_CPU=OFF   -DWITH_NVIDIA=ON   -DWITH_LINKED=ON   -DWITH_TORCH=OFF   -DGENERATE_PYTHON_BINDINGS=OFF   -DINFINI_OPS_OPS=/work/ops.json   -DCMAKE_CUDA_ARCHITECTURES=80
-- Operator selection: /work/ops.json
-- Resolved 1 linked operator source(s).
-- Configuring done
-- Generating done

cmake --build /work/cmake-build-linked --target infiniops --parallel 8
[100%] Built target infiniops

readelf -d libinfiniops.so
NEEDED: sampling.so
NEEDED: libtvm_ffi.so
RUNPATH: flashinfer sampling directory; tvm_ffi/lib; InfiniRT; CUDA

ldd -r libinfiniops.so
sampling.so => /opt/linked-site/flashinfer_jit_cache/jit_cache/sampling/sampling.so
libtvm_ffi.so => /opt/linked-site/tvm_ffi/lib/libtvm_ffi.so
(no missing libraries or unresolved symbols)

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	TVM-FFI integration build passed	Resolver tests passed; ELF dependencies and RUNPATH verified
Iluvatar	No	N/A - not affected	Hosted CI pending
MetaX	Yes	Hosted CI pending	Existing linked Torch source partition is shared
Cambricon	Yes	Hosted CI pending	Existing linked Torch source partition is shared
Moore	Yes	Hosted CI pending	Existing linked Torch source partition is shared
Ascend	No	N/A - not affected	Hosted CI pending

Benchmark / Performance Impact

N/A. This PR changes configure-time resolution and build wiring only.

Notes for Reviewers

• link_libraries entries are retained independently of manifest sort order, so --as-needed cannot discard a declared dependency.
• include_glob is applied for any linked library that declares one.
• WITH_TORCH=OFF disables generated ATen implementations, but the existing WITH_LINKED setup still requires an installed torch package for its shared Python/C++ ABI configuration; the documentation now states this explicitly.
• External DSOs remain environment-bound dependencies and are not bundled into the wheel.
• The follow-up to this PR will reduce feat(linked): add FlashInfer sampling provider #930 to the FlashInfer slot-16 provider only.
• The validation target was infiniops. The default all-target build also reached and built infiniops, then unrelated examples failed against the reused older InfiniRT prefix because its runtime copy API no longer matches current examples.