[HRX] Export the allocation in hrx_buffer_get_device_ptr

[Geramy]

hrx_buffer_get_device_ptr only ever tried a host mapping, so it returned HRX_STATUS_UNAVAILABLE for any device-local buffer. The comment beside it already said what was missing:

// For local-task (CPU) devices, the device pointer is available via mapping.
// For real GPU devices, this would use iree_hal_buffer_export.
// For now, map the buffer to get a usable pointer.

The consequence is that nothing needing a device address is reachable through the C API:

• a peer copy has no source address to name;
• hsa_amd_agents_allow_access has no allocation to grant, so device pools stay DISALLOWED_BY_DEFAULT;
• a caller wanting hsa_amd_memory_async_copy cannot describe either end, and falls back to the HAL's blit kernels.

On an 8x gfx1201 box that last one is the difference between a copy engine and the shader cores: HIP reaches ~48 GB/s H2D and ~56 GB/s D2H on the same link where a blit-kernel copy measures about 8.

This exports the buffer through the owning device's allocator with IREE_HAL_EXTERNAL_BUFFER_TYPE_DEVICE_ALLOCATION and returns the device pointer. The existing paths are untouched and still tried first, so an already-mapped buffer and a local-task device behave exactly as before, and a failed export falls through to the mapping attempt rather than becoming an error.

Tested on gfx1201 (Radeon AI PRO R9700, ROCm 7.13): device-local buffers now answer with an address, hsa_amd_memory_async_copy accepts both ends and completes, and the previous behaviour is unchanged for host-visible buffers.

[Geramy]

@jimw567 Please take these PRs as suggestions made during creation of an HRX backed engine, if you would like them in hrx let me know what would need to follow to get it merged please.