Resize Image/Mask loses the batch layout when Lanczos resizes a mask to one row

### Custom Node Testing

- [ ] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

Resizing two masks to 16×1 should return a mask tensor shaped `(2, 1, 16)`. Passing it to MaskToImage should produce two images shaped `(2, 1, 16, 3)`.

### Actual Behavior

ResizeImageMask returns `(2, 16)` with Lanczos. MaskToImage then interprets this as one mask and returns `(1, 2, 16, 3)`, merging the two batch items into one two-row image.

The grayscale path in `comfy.utils.lanczos` returns `[B, H, W]` instead of restoring `[B, 1, H, W]`. The resize node then squeezes dimension 1, which removes the height when it is 1.

### Steps to Reproduce

Run this from the ComfyUI checkout with its Python dependencies installed. It calls only the built-in nodes on CPU. Reproduced on current master `6f3895ed8d1e0f1d5fad11b48b3f159f7de1cc15` (0.35.0).

```python
from comfy.cli_args import args
args.cpu = True

import torch
from comfy_extras.nodes_post_processing import ResizeImageMaskNode, ResizeType
from comfy_extras.nodes_mask import MaskToImage

mask = torch.ones(2, 8, 16)
resized = ResizeImageMaskNode.execute(mask, "lanczos", {
"resize_type": ResizeType.SCALE_DIMENSIONS,
"width": 16,
"height": 1,
"crop": "disabled",
}).result[0]
print("resized mask:", tuple(resized.shape))
print("converted image:", tuple(MaskToImage.execute(resized).result[0].shape))
```

### Debug Logs

```powershell
torch/jit/_script.py:1491: FutureWarning: `torch.jit.script` is deprecated. Please switch to `torch.compile` or `torch.export`.
warnings.warn(
resized mask: (2, 16)
converted image: (1, 2, 16, 3)
```

### Other

Windows, Python 3.12.14, PyTorch 2.14.0+cpu, Pillow 12.3.0.

Related history: #12679 fixed transposed width/height for grayscale Lanczos output, but the channel axis is still missing. This report covers the remaining shape loss.

Start with the grayscale path in comfy.utils.lanczos and ResizeImageMaskNode in comfy_extras/nodes_post_processing.py, using the supplied CPU reproduction as the first check. The fix is complete when resizing a two-item mask to 16×1 preserves the channel dimension and MaskToImage returns shapes (2, 1, 16) and (2, 1, 16, 3).

Written by the indexing model from the issue text.