[Feature request] HW compatibility tags for workflow templates and nodes

## Problem
Users on Apple Silicon (MPS), low-VRAM GPUs, or AMD (ROCm) routinely waste time downloading models and running workflows that can't efficiently execute on their hardware. There's no way to filter templates by supported backend.

## Solution
Add optional `supported_backends` field to:

1. Workflow templates (in Comfy Registry):
```json
{
"supported_backends": ["cuda"],
"min_vram": 12,
"recommended_vram": 24
}
```

2. Node metadata (in node definitions):
```python
SUPPORTED_BACKENDS = ["cuda", "mps", "cpu"]
```

## Frontend changes
- Filter templates by backend in Workflows → Templates
- Show compatibility badge when template requires CUDA but user runs MPS
- Node palette grays out nodes incompatible with current backend

## Related
Issue #10717 (OOM on official templates), #1305 (node-level metadata)

Start with Comfy Registry workflow templates and node definitions, then trace the Workflows → Templates view and node palette entry points. Determine how backend and VRAM capabilities are detected for CUDA, MPS, ROCm, and CPU. Done means templates can be filtered, incompatible badges appear, and incompatible nodes are visibly marked as described.

Written by the indexing model from the issue text.