Add APU unified memory (GTT) support for Strix Halo / gfx1151

[DasPauluteli]

Summary

• Adds real-time tracking of the GTT (Graphics Translation Table) pool — the system RAM dynamically allocated to the GPU via AMD's TTM kernel subsystem
• Automatically detects APUs (Strix Halo, Phoenix, Hawk Point, Rembrandt, etc.) by device name or size heuristic (GTT ≥ 4× VRAM and VRAM < 8 GB)
• Shows a purple "Unified RAM (GTT)" progress bar in the UI when an APU is detected
• Relabels the VRAM bar as "VRAM (reserved pool)" on APUs to avoid the misleading impression that VRAM is the total available GPU memory
• Adds is_apu, gtt_used, gtt_total, gtt_used_percent, and device_name to the WebSocket payload
• Updates README with an APU/unified memory section explaining the GTT/TTM architecture

Test plan

• Verify monitor still works correctly on a discrete AMD GPU (VRAM bar shown, GTT bar hidden)
• Verify GTT bar appears and VRAM is relabeled on an APU (e.g. Strix Halo / gfx1151)
• Check rocm-smi --showmeminfo gtt --json returns expected keys on your system
• Confirm APU detection heuristic triggers correctly (check is_apu in WebSocket payload)

🤖 Generated with Claude Code

[DasPauluteli]

Preview:

[DasPauluteli]

I will continue the tests using an AMD Radeon RX 9060 XT (gfx1200) tomorrow.

[DasPauluteli]

Tested today on the RX 9060 XT (gfx1200) with the latest TheRock PyTorch nightly builds and PyTorch nightly builds. Both the discrete GPU path (VRAM bar, no GTT bar) and APU path work correctly.

Test plan

• Verify monitor still works correctly on a discrete AMD GPU (VRAM bar shown, GTT bar hidden)
• Verify GTT bar appears and VRAM is relabeled on an APU (e.g. Strix Halo / gfx1151)
• Check rocm-smi --showmeminfo gtt --json returns expected keys on your system
• Confirm APU detection heuristic triggers correctly (check is_apu in WebSocket payload)

All test plan items pass. Marking as ready to merge

[DasPauluteli]

I have the preview screenshot to the repo and README.md.

Preview: