result type of `argsort`: make it I64?

[foldl]

At present, result type of argsort is tensor of I32. It is better to be I64.

llama.cpp/ggml/src/ggml-cpu/ops.cpp

Lines 8648 to 8650 in b8e09f0

	for (int64_t j = 0; j < ne0; j++) {
	dst_data[j] = j;
	}

---

I am trying to use ggml_set_rows, and finding that it uses indices of I64. This is the motivation of creating this issue.

[JohannesGaessler]

Unless there is a specific use case that specifically needs the 64 bit range the indices should stay 32 bit, as this is better for performance and memory use. But it would make sense to add an assert for the max value that could be written as int32_t.

[ggerganov]

I think we can extend set_rows to accept both I32 and I64. There are certainly use cases for set_rows that require I64 (i.e. when writing cache values in very large KV cache buffers), so it has to continue to work with I64. But there are also use cases like in #16063 (comment) where the input indices are not I64 and it's not great to perform explicit cast.

argsort by itself seems to be OK to continue to work with I32 because we don't expect to perform this operator on billions of elements. So I don't think there is a use case to change argsort for now.