amdgpu: count GTT in per-process memory usage

[rborkow]

Problem

Per-process memory on amdgpu only reads the drm-memory-vram fdinfo key.
On integrated GPUs, nearly all allocations live in the GTT heap, so the
process list reports ~0 memory while the device-level metrics show
gigabytes in use. This resolves the existing TODO at the parse site
(// TODO: do we count "gtt mem" too?).

Measured on a Ryzen AI 9 HX 370 / Radeon 890M (gfx1150, kernel 6.18),
with a 24 GB LLM loaded by ollama's llama-server:

$ grep drm-memory /proc/<pid>/fdinfo/7
drm-memory-cpu:         0 KiB
drm-memory-gtt:  25258000 KiB   <- the model
drm-memory-vram:     5348 KiB   <- all nvtop counted

nvtop showed the device at 24.3/48.0 GiB used, but attributed only
5 MiB to the process — and gpu_mem_usage read 0%.

Change

Count drm-memory-gtt (and the legacy gtt mem key) together with
VRAM in parse_drm_fdinfo_amd, accumulating both into
gpu_memory_usage. This matches the device-level accounting, which
already sums both heaps via amdgpu_query_info (total_memory /
used_memory = vram + gtt). With numerator and denominator now
measuring the same thing, gpu_mem_usage percentages become coherent:
the process above reports 24.1 GiB / 50%, agreeing with the device
total.

Behavior change on discrete GPUs

A process's reported memory now includes its GTT usage (host RAM mapped
for the GPU) on dGPUs as well. I'd argue this is consistent — the
device-level metric already includes GTT, and both heaps are GPU-mapped
memory — but if you'd prefer to keep dGPU semantics unchanged, I'm happy
to rework this to add GTT only when static_info->integrated_graphics
is set.

Testing

Verified on the Radeon 890M above (before: ~0 per process; after:
per-process totals match device totals under a 24 GB workload).