refactor(triton): simplify JIT backend architecture

[voltjia]

Summary

• Separate backend-neutral Triton JIT orchestration from the NVIDIA runtime implementation through the JitBackend<device_type> template boundary.
• Split backend, compiler, cache, and JIT config responsibilities into focused modules; remove wrappers that had no independent semantics.
• Use structured backend metadata and one shared Triton config parser in generated Python bindings.
• Tighten kernel and auto-tuning cache identity, argument ordering, device/stream handling, artifact publication, and failure recovery.
• Add focused wrapper-generation, compiler, configured Add, empty-tensor, and polymorphic context coverage.

Motivation

This is a stacked cleanup of #800 focused on architecture, naming, modularity,
readability, and maintainability while preserving the template boundary needed
for future Triton backends.

The branch also carries the corresponding #919 changes because #800 has not
yet been rebased onto the master commit containing that merged PR.

Related: #800, #919

Type of Change

• refactor - code restructuring without an intended public behavior change
• Breaking change

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

Validated commit de9694aef9aef5bfb1c2705536b3d00592d6e318 on
accelerator-dev/nvidia:latest (8 x A100-SXM4-80GB, CUDA 13.1.80,
Python 3.12.3, Triton 3.5.1) against InfiniRT
0cdbb16967e15f2e055dea1ec9641617bf3b6cf6.

InfiniRT WITH_NVIDIA build/install: PASS
InfiniOps WITH_NVIDIA + WITH_TRITON + Python bindings build/install: PASS
Ruff 0.15.22 check/format: PASS
clang-format 21.1.8 on changed C++ files: PASS
focused generator/compiler tests: 32 passed
focused Add Triton tests: 3 passed, 3 skipped, 266 deselected
compile config on cuda:1 with current cuda:0: PASS (device 0 -> 0 -> 0)
auto-tuning on cuda:1 with current cuda:0: PASS (device 0 -> 0 -> 0)

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Pass	Triton JIT build, configured Add, auto-tuning, and multi-device guard passed
Iluvatar	No	N/A
MetaX	No	N/A
Cambricon	No	N/A
Moore	No	N/A
Ascend	No	N/A

Benchmark / Performance Impact

N/A. This PR does not make a performance claim.

Notes for Reviewers

• feat(triton): add JIT backend with add operator  #800 comes from a fork, so GitHub cannot use its head directly as an upstream PR base. stack/pr-800-triton-backend-jit is an upstream mirror pinned to feat(triton): add JIT backend with add operator  #800 head 0c73a9e262291852ef7bccee67c65e0f85e555ac; this PR should be retargeted after feat(triton): add JIT backend with add operator  #800 is rebased or merged.
• feat!: support polymorphic handles and configs #919 is already merged. This branch contains its corresponding changes under pre-rebase commit SHAs; they should disappear when feat(triton): add JIT backend with add operator  #800 is rebased onto current master.
• Python calls with an explicit Triton config intentionally use uncached Make so derived config identity is not erased by the generic operator cache. Compiled kernels and auto-tuning results remain cached by their complete identities. Configless calls continue through Call.
• Core JIT code reports failures through status values and the Python C API. The config parser retains one localized pybind11::value_error boundary so invalid binding input is exposed as Python ValueError.
• The backend template layer is retained for future platforms; wrappers without their own semantics were removed.