Taichi for Infernux

Derived from Taichi: https://github.com/taichi-dev/taichi
Copyright (c) The Taichi Authors (2016- ). All Rights Reserved.
Licensed under the Apache License, Version 2.0. See LICENSE.
Original author and copyright notices are retained in the source.

Infernux maintains this fork:
https://github.com/ChenlizheMe/taichi_for_infernux
https://github.com/ChenlizheMe/Infernux

Infernux modifications begun in 2026 establish a JIT-only compiler build
policy, engine-wheel output layout, and fork-specific documentation/CI.
They also remove independent runtime/resource ownership from the active
compiler path, replace public-package imports with private engine imports,
and prune Python field/SNode/device-array authoring implementations.
Kernel specialization uses engine buffer descriptions instead of runtime
array placeholders; unused automatic gradient and class-kernel paths are removed.
Compiled kernel IR ownership is request-local rather than retained by global
Python/native kernel registries; generated-source entries are scoped to lowering.
Standalone function specialization/call APIs and their caches are removed in
favor of engine-declared inline helpers. Borrowed compiler views retain their
native owners; return annotations use the supported Python typing interfaces.
The native context now directly invokes the SPIR-V compiler; obsolete backend
interfaces, singleton finalization and global IR-counter resets are removed.
Compiler output no longer carries TIC serialization, hashing or launch handles.
Texture-format and device-capability descriptions no longer import upstream
device/runtime declarations. The old host-adapter transfer/command implementation
and its device API target are removed; backend policy is defined once for builds.
The unused LLVM kernel dispatcher and CUDA/DirectX driver includes are removed
from the active compiler sources.
The standalone C-API/AOT implementation, non-SPIR-V code generators, execution
runtimes, device backends and their legacy containers are removed from source.
The Python AOT/graph/module tools, field-tree builders, autodiff authoring
entry points and standalone command-line interface are removed as well.
Standalone Taichi packaging and upstream runtime-test build entries are removed;
the engine CMake install component remains the distribution path.
Compiler options are local to each context and default directly to Vulkan;
obsolete backend options and the mutable global configuration are removed.
The native binding is renamed _infernux_gpu_compiler. Unused SNode/texture,
mesh-field and external-call frontend bindings and mesh compatibility stubs
are removed; loading no longer configures an upstream runtime directory.
The ongoing integration narrows this dependency to Vulkan SPIR-V generation;
buffer ownership, device management and execution belong to Infernux.
This notice does not claim that the compiler-only migration is complete.
