refactor!: reduce InfiniCore to a component manifest

[voltjia]

Summary

• Reduce InfiniCore to a component manifest containing only its license, manifest documentation, contribution guidelines, and the InfiniRT, InfiniOps, and InfiniCCL gitlinks.
• Pin one compatible revision of each component while the cross-repository migration is reviewed.
• Remove the legacy runtime, operator, Python package, build system, tests, scripts, and repository workflows from this repository.
• Add a focused CONTRIBUTING.md for the shared component contribution conventions.

Current Component Revisions

• InfiniRT 0cdbb16967e15f2e055dea1ec9641617bf3b6cf6.
• InfiniOps f890afb4b2327f13ccdd3c1b6b0d49567c5fe00d, the merge revision of feat(linked): add FlashInfer sampling provider InfiniOps#930. It contains the exact-slot ops.json selector, the TVM-FFI linked-provider transport, and the FlashInfer TopKTopPSamplingFromLogits provider. Its ancestry does not contain the native sampling draft feat(nvidia): add top-k top-p sampling provider InfiniOps#922.
• InfiniCCL e8ccc0cb23ca5b1b1d63be29ba61ea031de72807, the current tip of the stacked fix(build): regenerate communication bridge for header changes InfiniCCL#57, issue/48/test: 重构rope测试脚本 #58, and issue/7: rmsnorm算子-昇腾 #59 series. Those PRs are still under review; this pin must be refreshed to their final merged revision before refactor!: reduce InfiniCore to a component manifest #1406 lands.

The linked FlashInfer provider keeps its DSOs external and requires flashinfer-jit-cache>=0.6.7,<0.7 and apache-tvm-ffi==0.1.10 in the NVIDIA consumer environment.

Pending Final Pin Refresh

The Moore end-to-end enablement is being reviewed separately:

• feat(moore): support paged flash attention prefill InfiniOps#819 and issue/961: fix metax init with preload #962 provide Moore variable-length and KV-cache flash attention.
• fix(moore): select MCCL bfloat16 architecture InfiniCCL#69 provides explicit/native Moore architecture selection required by BF16 collectives.

After those changes and InfiniCCL #57-#59 merge, this PR will update its InfiniOps and InfiniCCL gitlinks to the resulting merged revisions. The manifest will not permanently pin review-only stack heads.

Motivation

InfiniCore should coordinate compatible component revisions instead of maintaining duplicate runtime and operator implementations. Keeping one authoritative implementation in each component removes the legacy API surface and makes the repository boundary explicit.

Repository Surface

The resulting repository has exactly seven tracked paths and no build or packaging entry point of its own. README.md documents the manifest role and recursive checkout, while CONTRIBUTING.md documents the contribution contract.

Breaking Change

InfiniCore no longer provides the legacy in-tree runtime, operator, build, packaging, or Python APIs. Component implementations are provided exclusively by the pinned submodules.

Validation

• Verified git ls-files returns the intended seven paths.
• Verified all three component entries use gitlink mode 160000 and resolve to the revisions listed above.
• Verified git submodule sync --recursive and git submodule update --init --recursive check out all three exact component revisions, including the pinned nested InfiniOps .ci revision.
• Verified the Core range passes git diff --check; the final commit changes only the InfiniOps gitlink while the validated InfiniCCL and InfiniRT pins remain unchanged.
• At the pinned InfiniOps revision, the manual NVIDIA validation recorded by feat(linked): add FlashInfer sampling provider InfiniOps#930 passed 53 resolver/schema tests, a Release CUDA and Python build for SM80, and 17 FlashInfer slot 16 GPU tests. ldd -r reported no unresolved dependencies, and descriptor-cache stress retained 0 MiB of device memory after cleanup.

InfiniCore intentionally has no build or packaging entry point. End-to-end consumer build, packaging, graph execution, and model-command validation are owned by InfiniTensor/InfiniLM#506.

Supersedes #1405.

Closes #1373