feat(gpu): live usage/VRAM/temperature for AMD on Linux via amdgpu sysfs

[SVNKVN9]

Mirror the NVML poller pattern for AMD cards: a 2 s background thread reads /sys/class/drm/card*/device/ — gpu_busy_percent, mem_info_vram_, and hwmon/temp_input (preferring the sensor labeled "edge" to match NVIDIA's GpuTemp). No new dependency; gated on cfg(target_os = "linux").

To support vendor-specific dispatch beyond is_nvidia, GpuRow now carries vendor_id + device_id (PCI 16-bit values from wgpu::AdapterInfo). AMD samples match wgpu slots by device_id, which is unambiguous unlike the substring name match used for NVML.

UI cleanup that came out of testing on RX 5x0 (8 GiB):

• shrink label column 14→6 ("temperature" → "temp") so the bar isn't squeezed to nothing on narrow panels
• compact "used / total " formatter — both numbers share the unit of total, so "1.6 / 8.0 GB" instead of "1.59 GB / 8.00 GB" (~10 chars saved)
• recompute bar_w against the vram line's worst-case trailing width; the old constant assumed every line ended in just "100.0%"
• per-vendor unsupported_reason text (NVIDIA / AMD / Intel / other)

README updated to reflect AMD-on-Linux now reports live stats.

[mantvmass]

ok