amdgpu: fall back to hwmon when gpu_metrics v1 is too old

[DvDlVs]

Problem

On some AMD GPUs - notably older Vega 20 / datacenter (Instinct) cards such as the Radeon Instinct MI50 / Radeon Pro VII (device id 0x66a1) - MangoHud shows GPU load at 0% and power at 0W permanently.

Root cause

These cards expose a gpu_metrics sysfs file in the old v1 format that is smaller than gpu_metrics_v1_3 (the smallest version MangoHud knows how to parse). On this hardware:

• The constructor only checked that the gpu_metrics file existed and was readable to set gpu_metrics_is_valid = true.
• In get_instant_metrics (src/amdgpu.cpp), reading a v1 file that is too small returns early without populating the fields, leaving gpu_load_percent and average_gfx_power_w at 0.
• Because gpu_metrics_is_valid was still true, metrics_polling_thread then overwrote the correct hwmon values (gpu_busy_percent and power1_input) with those zeros.

Fix

In the AMDGPU constructor, read the metrics header and validate structure_size against the largest version we can actually parse. If the version is not supported (e.g. old v1 smaller than v1_3), mark gpu_metrics_is_valid = false so MangoHud falls back to the working hwmon sysfs nodes (gpu_busy_percent for load and power1_input for power).

Verified on an Instinct MI50: GPU load and power now report real values instead of 0.

Notes

• Only reads the first 4 bytes of the header (no behavior change for GPUs with a supported gpu_metrics version).
• Falls back gracefully - existing supported GPUs are unaffected.