[nixlbench] add worker part for device api (#2015)

## What?
Integrate Device API into nixlbench so transfers can be initiated from a
CUDA kernel instead of the host path.
Adds --use_device_api to enable GPU-kernel-initiated PUT transfers. When
enabled, --num_threads is repurposed as the CUDA kernel block size, and
the internal CPU thread count is forced to 1.
This builds on the kernel launch support landed in #1761.

## Why?
nixlbench could not exercise Device API end-to-end now. Wire the
existing device kernels into the worker so we can benchmark and validate
GPU-initiated WRITE transfers under a supported configuration.

## How?
- Add --use_device_api and validate required options at config parse
time (UCX, VRAM↔VRAM, --enable_pt, pairwise, SG, WRITE,
pipeline_depth=1, and a build with HAVE_UCX_GPU_DEVICE_API).

- In the NIXL worker, prepare local/remote nixlMemViews,
allocate/exchange a VRAM completion counter, and launch
nixlbenchLaunchDevicePut for the transfer path.

- On the target, wait on the device completion/error counters instead of
host-side request completion.

- Keep the existing host transfer path unchanged when --use_device_api
is not set.


<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Added an optional NIXL Device API mode for GPU-based transfer
benchmarking.
* Added GPU-resident completion tracking and automatic monitoring for
warmup and timed transfers.
* Added GPU memory preparation and cleanup for local and remote transfer
views.
* Added validation and status reporting for device API settings,
including CUDA block size and supported transfer configurations.

* **Bug Fixes**
* Device transfers no longer require remote completion signaling to be
explicitly enabled.
* Completion and error signaling now consistently use the remote counter
region.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->