feat!: align runtime API and add runtime dispatch

[voltjia]

Summary

• Move CUDA Runtime API-shaped public wrappers under infini::rt::runtime, keeping infini::rt free for InfiniRT-specific APIs.
• Add top-level infini::rt::set_runtime_device_type and infini::rt::runtime_device_type for runtime dispatch between enabled backends.
• Normalize runtime memcpy-kind constants to kMemcpy... names and remove the non-k aliases from the public API surface.
• Keep the generated runtime API declaration list in scripts/generate_public_headers.py for now, instead of parsing CUDA docs or headers.
• Constrain TensorView's tensor-like constructor and add focused test_tensor_view coverage.
• Update README examples for the new infini::rt::runtime namespace and runtime-device selector APIs.

Motivation

The latest interface guide requires the lowest-level C++ runtime API to match CUDA Runtime API signatures after only Google C++ Style naming conversion. Keeping those CUDA-shaped APIs under infini::rt::runtime makes that contract explicit, while the outer infini::rt namespace can host InfiniRT-specific dispatch APIs without worrying about current or future CUDA symbol conflicts.

Companion InfiniOps PR: InfiniTensor/InfiniOps#787

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

# Host validation on ssh nvidia, outside Docker
export PATH=/usr/local/cuda/bin:$HOME/.local/bin:$PATH
cmake -S /tmp/infinirt-host-bab8/src -B /tmp/infinirt-host-bab8/build \
  -DWITH_CPU=ON -DWITH_NVIDIA=ON -DINFINI_RT_BUILD_TESTING=ON \
  -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX=/tmp/infinirt-host-bab8/prefix
cmake --build /tmp/infinirt-host-bab8/build -j8
ctest --test-dir /tmp/infinirt-host-bab8/build --output-on-failure
cmake --install /tmp/infinirt-host-bab8/build

100% tests passed, 0 tests failed out of 8

# accelerator-dev/nvidia:latest, with companion InfiniOps branch
InfiniOps no-Torch InfiniLM operator subset wheel: built and installed successfully.
InfiniOps representative PyTorch wrapper smoke (`abs,clamp,exp`): built and installed successfully.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	yes	smoke passed	Host CPU+NVIDIA CMake build, CTest 8/8, install, and install-consumer passed. Companion InfiniOps no-Torch and representative PyTorch smoke builds passed in accelerator-dev/nvidia:latest.
Iluvatar	yes	N/A	Not tested: no Iluvatar hardware/toolkit in the validation environment.
MetaX	yes	N/A	Not tested: no MetaX hardware/toolkit in the validation environment.
Cambricon	yes	N/A	Not tested: no Cambricon hardware/toolkit in the validation environment.
Moore	yes	N/A	Not tested: no Moore hardware/toolkit in the validation environment.
Ascend	yes	N/A	Not tested: no Ascend hardware/toolkit in the validation environment.

Full `pytest` output (optional)

N/A

Benchmark / Performance Impact

N/A. This PR changes runtime API shape and dispatch plumbing; no performance-sensitive kernels are changed.

Notes for Reviewers

• The lowest-level CUDA-shaped APIs are now infini::rt::runtime::*; the outer infini::rt namespace currently only adds runtime-device selector APIs.
• The generated public header declarations are still driven by a short explicit list in the generator script. Parsing CUDA docs or cuda_runtime.h is intentionally left out of this PR to keep scope contained.
• Full InfiniCore/InfiniLM inference validation was attempted, but the remote environment repeatedly failed while downloading xmake's Boost dependency from GitHub. The failure occurred before InfiniCore compilation and appears network-related. The relevant InfiniRT and InfiniOps compatibility checks listed above passed.

BREAKING CHANGE: CUDA-shaped runtime APIs move to infini::rt::runtime, and non-k runtime memcpy-kind aliases are removed from the normalized runtime API surface.