build: consume installed InfiniOps package

[voltjia]

Summary

• consume the installed InfiniOps::infiniops CMake package through xmake instead of configuring, building, and installing the InfiniOps source tree from InfiniCore
• resolve the InfiniOps prefix from --infiniops-root, INFINI_OPS_ROOT, INFINI_ROOT, or ~/.infini
• retain INFINI_OPS_OPS only for selecting InfiniCore adapters and linked-attention defines; InfiniOps itself is built independently from the same ops.json
• stage the installed InfiniOps runtime and give its standalone InfiniRT dependency a private SONAME so it can coexist with InfiniCore's current bundled InfiniRT

Impact

InfiniRT and InfiniOps must be installed before configuring InfiniCore. --infiniops-root now points to an install prefix containing InfiniOpsConfig.cmake and libinfiniops.so; InfiniCore no longer owns an InfiniOps CMake build or requires the InfiniOps source tree for the NVIDIA build.

This PR is stacked on the adapter-filtering fix and requires the CMake package export merged in InfiniOps #948.

Validation

• git diff --check
• validated with InfiniRT 6b256e0ab085, InfiniOps 903ef9e026a8, and this branch at 6e712656840a5569f3e38af80426ef6d1006351b
• built and installed InfiniRT and InfiniOps independently into a fresh prefix in accelerator-dev/nvidia:latest
• configured InfiniCore with an empty submodules/InfiniOps, then built and installed infinicore_cpp_api and _infinicore
• confirmed the installed InfiniCore libinfinirt.so remained its bundled runtime while libinfiniops.so resolved libinfiniops_infinirt.so; ldd reported no missing dependency
• imported the freshly installed Python package without manual library repair
• rebuilt InfiniLM against the independent prefix and ran Qwen3-0.6B inference on NVIDIA with greedy decoding; prompt hello produced <think>\nOkay, the user just said \"hello\". I need to respond appropriately