Raise proper error when deconding 16-bits jpegs

[cyyever]

After some investigation, I found that #3613 failed because the image is 16 bit. However our decoding code assumes the image is 8-bit. So our allocated tensor size is too small to hold the decoded data and leads to memory overflow.

However, I can't find an easy way to decode 16-bit images without breaking the current decoding logic(we would have to change too many code places). So I simply issue an error.

Note: In order to support 16-bit png images, we should allocate the tensor with torch::kI16 and change

auto datap = data.accessor<unsigned char, 1>().data();

and

auto ptr = tensor.accessor<uint8_t, 3>().data();

and the logic of read_callback. And although #4051 didn't fix this bug, I still think it's useful..

[NicolasHug]

Thanks @cyyever , LGTM!

I can confirm that the offending 16-bits jpeg now fails with the more informative error message rather than a segfault.

@fmassa should we open an issue to keep track of the 16-bits support, or is this not something we want?

[cyyever]

And there may be 32-bit or 64-bit png images. And I'm not sure the case of jpg.

[NicolasHug]

Thanks @cyyever !

[fmassa]

Thanks for the fix @cyyever !