[CORE] Staging Object Tying Feature

[tqchen]

Summary

Object tying remains compiled but is deliberately staged dormant: canonical-wrapper detection returns false, while C++, Python-defined dataclasses, and Rust objects continue to use matching legacy allocation and deletion paths.

This staging has three goals:

• let downstream projects upgrade without making existing generated caches depend on the new allocator contract;
• keep the compatibility change compact by changing only the activation/allocation boundary instead of removing the object-tying implementation;
• preserve the custom allocator API and a clean path for a later coordinated object-tying activation.

Legacy wrapper identity and reference-count expectations are restored, while activation-only tying tests remain as skipped coverage for the later rollout.

Testing

• C++: 465 enabled tests passed (2 disabled).
• Python: 2,381 passed, 78 skipped, 2 expected failures.
• Rust: 138 unit/integration tests and 3 doctests passed; formatting is clean.
• Pre-commit: all changed-file checks passed, including ruff, ty, and clang-format.
• Order-balanced microbenchmarks place the dormant path about 1.3-7.9% behind active tying and 4.7-12.5% ahead of a full revert.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[tqchen]

Post-publication FlashInfer confirmation against published head 89031a1c:

• Rebuilt the real generated silu_and_mul CUDA source against the published headers for H200 (sm_90a). The resulting DSO has SHA-256 8c9e044ce5d239d946274918bd589e9fbdc6dbdb3b433b448facf2c329db0b52.
• The DSO has no undefined TVMFFIGetCustomAllocator or TVMFFISetCustomAllocator reference.
• With LD_BIND_NOW=1, isolated released tvm-ffi 0.1.11 and 0.1.12 runtimes both load the DSO and produce the expected argument-count TypeError; this exercises cache-side error-object construction without a symbol-resolution failure.
• Both released runtimes also complete a valid FP16 silu_and_mul launch on H200, synchronize successfully, and match the Torch reference at rtol=1e-3, atol=1e-3.
• In the reverse direction, the pre-allocator-change owned-Shape cache still loads, returns [2, 3], and destroys cleanly with the dormant runtime.

The detailed compatibility investigation and benchmark matrix remain on the task issue.