[AMDGPU] Embed prebuilt AMDGPU device binaries (#24498)

The AMDGPU HAL runtime needs a small builtin device library for blit,
dispatch, and timestamp kernels. Building that library through the
normal runtime target required an AMDGPU-capable clang from the
third_party LLVM checkout, making accidental runtime rebuilds extremely
expensive and tying the runtime to the monorepo compiler checkout.

This adds a device/binaries package that embeds checked-in HSACO code
objects for the default generic support set: gfx9-generic, gfx90a,
gfx9-4-generic, gfx10-1-generic, gfx10-3-generic, gfx11-generic, and
gfx12-generic. Default builds consume those blobs and do not require
ROCm or LLVM producer tools.

Default/prebuilt scenarios:
```
  iree-bazel-build //runtime/src/iree/hal/drivers/amdgpu/device/binaries:toc

  cmake -DIREE_HAL_AMDGPU_DEVICE_BINARY_BUILD_MODE=prebuilt ...
```

Live source-build scenarios for HAL developers:
```
  iree-bazel-build \
      --config=amdgpu_device_binaries_source_rocm \
      --repo_env=IREE_HAL_AMDGPU_DEVICE_TOOLCHAIN_ROCM_PATH=/path/to/rocm \
      //runtime/src/iree/hal/drivers/amdgpu/device/binaries:toc

  iree-bazel-build \
      --config=amdgpu_device_binaries_source_llvm_project \
      //runtime/src/iree/hal/drivers/amdgpu/device/binaries:toc

  cmake \
      -DIREE_HAL_AMDGPU_DEVICE_BINARY_BUILD_MODE=source \
      -DIREE_HAL_AMDGPU_DEVICE_TOOLCHAIN=rocm \
      -DIREE_HAL_AMDGPU_DEVICE_TOOLCHAIN_ROCM_PATH=/path/to/rocm ...
```

Regenerating the checked-in set is explicit:
```
  python build_tools/scripts/amdgpu_device_binaries.py \
    --output-dir runtime/src/iree/hal/drivers/amdgpu/device/binaries/prebuilt \
    --rocm-path /path/to/rocm \
    --targets gfx9-generic,gfx90a,gfx9-4-generic,gfx10-1-generic,gfx10-3-generic,gfx11-generic,gfx12-generic
```

The target map is authored once in Python and generates the Bazel,
CMake, and runtime lookup fragments. That keeps the ROCm/TheRock family
vocabulary, exact target map, and generic code-object fallback table
aligned without adding a manifest whose inventory, history, and hashes
would duplicate Git.

The pragmatic tradeoff is accepting checked-in binaries for code that
changes rarely. The current seven HSACO files are about 151 KiB raw,
about 40 KiB as loose Git blobs, and about 27 KiB in a Git pack, so the
repository cost is a few tens of KiB instead of a mandatory LLVM/ROCm
dependency on every runtime build.

This does not add general AMDGPU fatbin support. The builtin runtime
library now has a TOC of target-specific blobs and runtime selection;
CTS, samples, and other normal HAL executables still use the existing
ROCm test target plumbing until the AMDGPU executable format grows real
multi-variant support.