Add ggml_roll

[Acly]

This adds ggml_roll (torch.roll) as a new operation. It shifts tensor elements with wrap behavior.

It is used by eg. SWIN transformer to implement shift for its window attention.

[ggerganov]

Should we change the signature like so:

Suggested change

	// Move tensor elements by an offset given for each dimension. Elements that
	// are shifted beyond the last position are wrapped around to the beginning.
	GGML_API struct ggml_tensor * ggml_roll(
	struct ggml_context * ctx,
	struct ggml_tensor * a,
	int shift0,
	int shift1,
	int shift2,
	int shift3);
	// Move tensor elements by an offset given for each dimension. Elements that
	// are shifted beyond the last position are wrapped around to the beginning.
	// If tensor b is provided, the shifts become: shiftX' = shiftX + b[X]
	GGML_API struct ggml_tensor * ggml_roll(
	struct ggml_context * ctx,
	struct ggml_tensor * a,
	struct ggml_tensor * b, // optional I64 [GGML_MAX_DIMS]
	int shift0,
	int shift1,
	int shift2,
	int shift3);

This in theory would allow to pass the shifts as dynamic graph input data, which makes the graph static in terms of nodes (and op params). Could be useful for graph reuse.

[slaren]

I wouldn't bother with this unless there is some use case that you are thinking about, otherwise this is likely to end increasing complexity for no real reason. If you decide to add this, for clarity I would recommend adding a different function that takes only the tensor, rather than trying to do everything with a single function and optional parameters.

[ggerganov]

Ok, let's merge as is for now.