Claude review: drm/v3d: Refcount v3d_stats

[Claude Code Review Bot <claude-review@example.com>]

Patch Review

This is the core change of the series. The design is sound: `v3d_stats` gets a `kref` and is heap-allocated via `v3d_stats_alloc()`.

The error handling in `v3d_open` is correct — the two-label approach (`err_sched` / `err_stats`) properly handles the case where stats were allocated but `drm_sched_entity_init` failed:

```c
err_sched:
	v3d_stats_put(v3d_priv->stats[i]);
err_stats:
	for (i--; i >= 0; i--) {
		drm_sched_entity_destroy(&v3d_priv->sched_entity[i]);
		v3d_stats_put(v3d_priv->stats[i]);
	}
```

The `v3d_gem_init` error handling is also correctly reworked — previous bare `return -ENOMEM` / `return ret` paths that leaked stats are now routed through proper cleanup labels. The reordering that separates `drm_mm_takedown` from the DMA alloc failure path (`err_dma_alloc` vs `err_sched`) is correct.

One observation: `v3d_stats_release` and `v3d_stats_alloc` are placed in `v3d_sched.c` rather than, say, `v3d_drv.c` where `v3d_get_stats` lives. This is fine since `v3d_sched.c` is the primary consumer, but worth noting for future reference.

---
Generated by Claude Code Patch Reviewer