False positive "GPU/accelerator not supported" conflict for registry packs declaring "GPU :: NVIDIA CUDA"

### Description
The Custom Node Manager flags a false positive "accelerator" conflict for any registry node pack whose version metadata declares `supported_accelerators: ["GPU :: NVIDIA CUDA"]`, on machines with a CUDA GPU.

The package info panel shows:
> GPU/accelerator not supported (available: CUDA, required: GPU :: NVIDIA CUDA)

and the pack appears under the "Conflicting" category, even though the environment fully supports it.

### Steps to Reproduce
1. Install a registry pack whose node version declares `supported_accelerators: ["GPU :: NVIDIA CUDA"]` (e.g. `minimax-h3-audio-t8`).
2. Open Custom Node Manager.
3. The pack appears under "存在冲突" (Conflicting).
4. Open its info panel: it shows the false warning above.

### Root cause
In `useConflictDetection`, `checkAcceleratorCompatibility` compares with an exact array element match:

```js
const n = getRegistryAccelerator(systemAccelerator); // "cuda" -> "CUDA"
return e.includes(n) ? null : { type: 'accelerator', ... };
```

`e` is the pack's `supported_accelerators` from the registry, which uses classifier-style values such as `"GPU :: NVIDIA CUDA"`, so `["GPU :: NVIDIA CUDA"].includes("CUDA")` is `false` on every CUDA machine. The same likely affects `"GPU :: Apple MPS"`/`Metal` and `"GPU :: AMD ROCm"`/`ROCm`.

### Suggested fix
Match after normalization, e.g.:
```js
return e.some(x => x.toLowerCase().includes(n.toLowerCase())) ? null : { type: 'accelerator', ... };
```

### Environment
- ComfyUI core: v0.35.1 (ComfyUI Desktop, Windows)
- comfyui_frontend_package: 1.51.10
- OS: Windows
- Accelerator: CUDA (NVIDIA GPU)
- Not specific to one custom node: any registry pack declaring `supported_accelerators: ["GPU :: NVIDIA CUDA"]` triggers it.
-

Start in useConflictDetection and inspect checkAcceleratorCompatibility, where registry supported_accelerators is compared with the normalized system accelerator. Reproduce with a pack declaring "GPU :: NVIDIA CUDA" on a CUDA environment, then verify the pack is no longer marked conflicting and the warning disappears; check the analogous Apple MPS/Metal and AMD ROCm mappings described in the issue.

Written by the indexing model from the issue text.