Introduce layered shared utility libraries (common_utils + miopen_utils)

[BradPepersAMD]

Summary

• Create two new shared utility libraries to eliminate circular dependencies between MIOpen driver, test, and library code
• Move 13 pure C++ utility headers from src/include/miopen/ to common_utils/
• Move 13 shared verification/test utility headers from test/ to miopen_utils/
• Eliminate all cross-includes between driver/ and test/ in both directions
• Remove 15 unnecessary includes where test files included driver headers without using them

Motivation

MIOpenDriver, tests, and the MIOpen library were tangled through circular includes:

• 51 driver files included test headers
• Test tensor_holder.hpp included driver/random.hpp (circular)
• 18 test files included driver headers (14 unnecessarily)

This prevents building components independently and makes the dependency graph fragile.

Architecture

Note: Both common_utils and miopen_utils are internal build-only libraries.
They are not installed, not exported via find_package(MIOpen), and are not part of MIOpen's
public API. They exist solely to give the build system explicit, enforced edges between
driver/, test/, and src/. Consumers outside this repository should never depend on them.

This PR is Phase 1 of an incremental migration. Some layer boundaries listed below are not
yet fully enforced in code (e.g., common_utils/random.hpp still depends on miopen/env.hpp).
Known violations are listed as Phase 2 follow-up items. This PR is fully functionally equivalent
to the state before it — no runtime changes, no new build failures.

Layer 1: common_utils (no dependencies)

Pure C++ utilities. Depends only on STL/system headers. No MIOpen, no HIP, no GPU.

Layer 2: MIOpen Library (depends on: common_utils)

The MIOpen library itself. Uses common_utils internally. Exports the public miopen.h C API.

Layer 3: miopen_utils (depends on: common_utils, MIOpen public API)

Shared verification/test utilities. Uses common_utils and miopen.h. Cannot use MIOpen internal headers.

Layer 4: MIOpenDriver + Tests (depends on: common_utils, miopen_utils, MIOpen public API)

End consumers. Use all three lower layers via public interfaces. Never include MIOpen internals or each other.

common_utils (Layer 1) -- 13 headers

Pure C++ utilities with zero MIOpen dependencies:

• ford.hpp, par_for.hpp -- parallel/sequential loop utilities
• each_args.hpp, functional.hpp -- variadic template / function composition
• float_equal.hpp -- ULP-based float comparison
• bfloat16.hpp -- bfloat16 type (config dependency replaced with CMake define)
• stringutils.hpp -- string utilities (MIOPEN_THROW replaced with std::runtime_error)
• rank.hpp, returns.hpp, type_name.hpp, algorithm.hpp -- metaprogramming
• reduce_common.hpp -- type conversion helpers
• random.hpp -- PRNG (moved from driver/, breaks circular dep)

miopen_utils (Layer 3) -- 14 headers

Shared verification utilities used by both driver and tests. Depends on common_utils AND the MIOpen public API:

• tensor_holder.hpp, verify.hpp -- core tensor container + numerical verification
• cpu_conv.hpp, cpu_bias.hpp, cpu_layernorm.hpp -- CPU reference implementations
• fusionHost.hpp, gemm.hpp, cpu_reduce_util.hpp, rnn_util.hpp -- more CPU refs
• random.hpp, serialize.hpp, network_data.hpp -- tensor init + serialization
• mloSoftmaxHost.hpp -- CPU softmax reference (moved from driver/)
• gpu_mem.hpp -- GPUMem forwarding shim (Phase 2: extract from driver.hpp)

Backward compatibility

All original file locations retain thin forwarding headers, so existing MIOpen library and test code continues to work without changes.

Forwarding shims keep existing include paths working within the build tree.
No public headers or installed paths change.

Cross-include status

Boundary	Before	After
driver/ to test/	54 includes	0
test/ to driver/	18 includes	0
miopen_utils/ to test/	n/a	0
miopen_utils/ to driver/	n/a	1 (gpu_mem.hpp shim, Phase 2)
common_utils/ to miopen/	n/a	1 (random.hpp to env.hpp, Phase 2)

Phase 2 (follow-up)

• Extract GPUMem class from driver.hpp into miopen_utils standalone header
• Remove random.hpp dependency on miopen/env.hpp
• Replace miopen::deref() and other internal symbol usage in driver
• Refactor tensor_holder.hpp to use public MIOpen API instead of internal TensorDescriptor

Test plan

• Build MIOpen library
• Build MIOpenDriver
• Build and run tests
• Verify no circular includes between driver/ and test/
• Verify common_utils/ has no miopen/ dependencies (except noted Phase 2 item)