refactor(runtime)!: adopt the modern Infini stack

[voltjia]

Summary - Move the runtime context, tensor layer, operator adapters, graph bridge, distributed wrappers, and Python bindings into InfiniLM so the inference repository owns its integration layer. - Build InfiniRT, a focused canonical InfiniOps closure, and InfiniCCL from the revisions pinned by InfiniCore. - Select exact InfiniOps implementation slots from a consumer-owned scripts/configs/infiniops_ops.json: native slot 0 where appropriate, generated Torch slot 8 for Argmax, linked slot 16 for FlashAttention and FlashInfer sampling. - Replace production *Infinilm operator APIs with canonical APIs and restrict the compiled modern model surface to supported dense non-quantized Qwen3. - Rebase the complete 20-commit migration onto current main 58b8c0cb. ## Motivation The inference stack previously depended on runtime and operator implementations duplicated across repository boundaries. This change makes InfiniRT, InfiniOps, and InfiniCCL authoritative while keeping inference-specific integration code with InfiniLM. Related to InfiniTensor/InfiniCore#1373. ## Component Contract - InfiniRT 0cdbb16967e15f2e055dea1ec9641617bf3b6cf6 - InfiniOps 94965ca433ea44b9cd4b8520a0f1d9687f81cfa9 - InfiniCCL e8ccc0cb23ca5b1b1d63be29ba61ea031de72807 InfiniTensor/InfiniOps#931 provides unified exact-slot selection through ops.json; stacked InfiniTensor/InfiniOps#930 provides TVM-FFI transport and FlashInfer TopKTopPSamplingFromLogits slot 16. Native sampling PR InfiniTensor/InfiniOps#922 is not in the selected InfiniOps ancestry. The FlashInfer linked DSOs remain external runtime dependencies under the existing linked-provider contract. The build environment must provide flashinfer-jit-cache>=0.6.7,<0.7 and apache-tvm-ffi==0.1.10. ## Validation ### Existing migration baseline The prior exact component combination passed a clean A100 wheel build, import, static and paged attention, eager and graph execution, batch 1/2, greedy/non-greedy sampling, TP1/TP2, PP2/TP1, and TP2+PP2. ### Final FlashInfer provider - InfiniOps Release/NDEBUG build passed without provider warnings. - Linked resolver tests passed 33/33. - FlashInfer slot 16 GPU tests passed 10/10, including both filter orders, host/CUDA indices, int32/int64 output, seed/offset, and a non-default-stream nonblocking sentinel. - InfiniLM range-diff confirms all 20 migration commits survived the rebase unchanged; git diff --check passes. - The final integrated InfiniLM wheel/Qwen smoke against the new component SHA is in progress; this PR remains draft until that run is recorded here. ## Supported Boundary - Validated platform: NVIDIA A100. - Compiled model: dense non-quantized Qwen3. - Unsupported quantization and linear-bias configurations are rejected before worker startup. - Paged attention uses the default 256-token block size; custom block sizes remain unvalidated. - Other model families and platforms remain outside this migration validation. ## Type of Change - [x] refactor - [x] fix - [x] test - [x] docs - [x] build / CI - [x] breaking change ## Landing Order 1. InfiniTensor/InfiniOps#931. 2. InfiniTensor/InfiniOps#930, rebased/retargeted after InfiniTensor/InfiniOps#931 merges. 3. InfiniTensor/InfiniCCL#57, InfiniTensor/InfiniCCL#58, and InfiniTensor/InfiniCCL#59. 4. InfiniTensor/InfiniCore#1406. 5. This PR.