[Runtime] Migrate from high-level iree_runtime API to low-level VM/HAL API

[sjain-stanford]

Summary

Fixes #82.
Fixes #15.

Replace the high-level iree_runtime_* wrapper layer with direct iree_vm_* and iree_hal_* APIs for precise lifetime control and device management.

Key type mapping

Removed (high-level)	Replaced with (low-level)
iree_runtime_instance_t	iree_vm_instance_t + global driver registry
iree_runtime_session_t	iree_vm_context_t with explicit HAL/bytecode module registration
iree_runtime_call_t	iree_vm_list_t + iree_vm_invoke

Major changes

• VM/HAL API migration (runtime.h, backend.h, handle.h, graph.h): Replace all iree_runtime_* types and calls with their low-level equivalents across instance creation, device creation, context setup, bytecode loading, and function invocation (addresses [Core] Move away from highlevel iree_runtime_* API #82).
• HAL driver registration (runtime.h): Extract driver registration into registerHalDriversOnce() using std::call_once, since the global driver registry persists across VM instance lifetimes and does not support re-registration.
• Function caching (runtime.h, graph.h): Cache resolved iree_vm_function_t in Graph::vmFunction_ during createVmContext() to avoid repeated lookups on each execute() call (addresses [Perf] Memoize runtime call initialization and reuse calls between Graph::execute invocations #15).
• Input list capacity pre-computation (runtime.h, graph.h): Pre-compute vmInputListCapacity_ during createVmContext() to avoid recomputing on every execute() call.
• Remove output list (runtime.h): Pass nullptr for outputs in iree_vm_invoke since compiled functions write results in-place (void return).

Bug fixes

• RAII cleanup for VM input list (runtime.h): Wrap iree_vm_list_t in IreeVmListUniquePtrType to prevent resource leaks on early-return error paths in execute().
• Early VM context ownership (runtime.h): Assign vmContext_ immediately after iree_vm_context_create so failures in HAL module creation or bytecode loading don't leak the raw context.
• Release HAL driver after AMDGPU device creation (runtime.h): createAMDGPUDevice() was not releasing the driver after device creation, leaking it on both success and failure paths. The device internally retains its own reference, so the caller's reference must be released separately (mirrors the pattern in createCPUDevice()).

Minor changes

• New RAII types (backend.h): Add IreeVmContextDeleter, IreeVmListDeleter and corresponding unique_ptr aliases.
• Include updates (backend.h, buffer.h, handle.h, runtime.h, test files): Replace <iree/runtime/api.h> with granular <iree/hal/api.h>, <iree/vm/api.h>, <iree/modules/hal/module.h>, etc.
• Redundant initializer cleanup (tensor_attributes.h, graph.h): Remove = std::nullopt on std::optional members that default-construct to empty.
• Comment marker normalization (tensor_attributes.h, graph_import.h, fusilli_plugin.cpp, test_fusilli_plugin_api.cpp, runtime.h): Normalize // C++ 20 → // C++20 across the codebase.
• Comment and documentation updates (handle.h, graph.h, runtime.h, test_graph.cpp): Update comments to reflect VM/HAL terminology.

No changes to CMakeLists.txt, samples, benchmarks, or public API.

🤖 Generated with Claude Code