[vello_hybrid] A single encoder with too many commands can permanently stall (Metal backend)

Nobody has claimed this yet.

### Context

This issue is split out from another broader performance investigation regarding Vello Hybrid. If Vello Hybrid has a lot of work to do, and pushes a lot of work onto the `CommandEncoder`, it is possible to permanently stall depending on the wgpu graphics backend.

### Repro

Run a very expensive scene (e.g. the complex compositing/mixing scene), on a large monitor using the Metal graphics backend. The scene will never render. Using `lldb` to inspect the process yields:

```
frame #3: 0x00000001a3b8c90c Metal`-[_MTLCommandBuffer initWithQueue:retainedReferences:synchronousDebugMode:] + 188
```

We get the same diagnostics as https://github.com/gfx-rs/wgpu/issues/1960.

As a quick validation, I wrote a proof-of-concept where Vello Hybrid owns the `CommandEncoder` and periodically flushes it. With this change the stall no longer reproduces.

Run the complex compositing/mixing scene with the Metal graphics backend on a large monitor and inspect the Vello Hybrid CommandEncoder ownership. Use the referenced wgpu diagnostics and the proof-of-concept periodic flush as investigation points. Done means the expensive scene renders without permanently stalling.

Written by the indexing model from the issue text.