feat: add Jetson Xavier unified memory detection

[hernandez42]

Summary

Add Xavier family markers to _NVIDIA_UNIFIED_MEMORY_MARKERS so whichllm correctly detects NVIDIA GPUs on Jetson Xavier platforms that use unified memory.

Changes

• File: src/whichllm/hardware/nvidia.py
• Added "JETSON AGX XAVIER", "JETSON XAVIER NX", and "XAVIER" to the unified memory marker tuple

Problem

Jetson AGX Xavier and Xavier NX devices were not detected as having a GPU because their hardware names weren't in the unified memory detection list. whichllm would report no GPU found even though the device has a Maxwell GPU with unified memory.

Solution

Extended the marker list to cover all Xavier variants, matching the existing pattern used for Orin (GB10, DGX SPARK).

Testing

On a Jetson AGX Xavier or Xavier NX device:

pip install whichllm
whichllm --detect

Should now show the GPU model and unified memory size instead of 'No GPU detected'.

[Andyyyy64]

really sorry this took me so long to get to, and thanks for the contribution.

Xavier does use unified system memory, so the direction makes sense. I found a few things we need to clear up before merging:

• ruff format --check . currently fails on src/whichllm/hardware/nvidia.py.
• the marker is only consulted after NVML or nvidia-smi returns a non-empty GPU name. so if the real failure is genuinely “No GPU detected” because neither path enumerates a GPU row, adding XAVIER here won’t fix that case. could you include the actual Xavier output or error from NVML / nvidia-smi, and add a regression test that reproduces that exact path?
• the architecture comments are mixed up: Jetson AGX Xavier uses a Volta GPU, not Maxwell, while GB10 / DGX Spark are Grace Blackwell systems rather than Jetson Orin. NVIDIA’s specs confirm both: Jetson AGX Xavier and DGX Spark.

once we have the real device name/output captured in a test and the comments/formatting corrected, this should be straightforward to re-review.