[HAL/AMDGPU] Name prepublished kernarg storage

Reusable AQL command buffers were materializing prepublished kernarg templates by asking the HAL allocator for DEVICE_LOCAL|HOST_VISIBLE memory and trusting the AMDGPU allocator to resolve that to a fine-grained host-coherent device pool. That was the right local outcome, but the contract was anonymous: a future allocator policy change could silently put template bytes behind a non-coherent mapping and make replay correctness depend on a late flush branch.

Introduce a small AQL prepublished-kernarg storage strategy and record the current device-fine host-coherent strategy on each physical device. Logical-device command-buffer creation passes the selected strategy into the command buffer after queue affinity has been normalized to one physical device, and recording only prepublishes reusable static dispatches when the strategy is enabled.

Finalization now requests DEVICE_LOCAL|HOST_VISIBLE|HOST_COHERENT explicitly and verifies that the returned AMDGPU buffer actually has those memory-type bits before copying templates. The old non-coherent flush fallback is gone because it was not a real strategy. Missing fine-grained device-local pools also get explicit allocator and physical-device diagnostics instead of an opaque pool lookup failure.

The low-level command-buffer tests now use a real heap allocator instead of constructing command buffers with a null allocator, preserving the production object invariant while keeping prepublish storage disabled for tests that only exercise recording or block processing. Add a host-queue integration check for the recorded device-fine strategy alongside the existing real prepublished-dispatch execution test.