Please add 'keep_channels' flag to make_grid

🚀 Feature

For images with 1 channel, it would be useful to tell make_grid to not convert grayscale images to RGB

Motivation

I just wanted to do a simple MNIST example but torchvision.utils.make_grid modified the data such that it became 3-dimensional RGB. That's cool for color images but I wish there were a simple way to keep the channels the same

Pitch

if the user passes "keep_channels=True" then the number of channels doesn't change

Alternatives

first i looked for other issues, and found others have the same problem
i tried using cv2 but it had some error,
i tried doing the y = 0.2989 * r + 0.5870 * g + 0.1140 * b math but it didn't look great,

so eventually i settled on converting the tensor to PIL,
converting the PIL to grayscale,
then converting the grayscale image back to a tensor,
which is a ton of extra code and compute,
when we could just not convert images to RGB in the first place!

Additional context

https://github.com/pytorch/vision/issues/863

Thank you for making torchvision

Start by locating the torchvision.utils.make_grid entry point and its existing tests, then trace where single-channel images are expanded to RGB. Add coverage for the proposed keep_channels option and verify that a one-channel input retains its channel count while existing color-image behavior remains unchanged.