CAGRA: Separate graph index sorting functionality from prune function

[enp1s0]

Changes

This PR separates the graph index sorting functionality from the CAGRA pruning function and creates a new function. (Related issue: #1446)

Unit test

I have included a new unit test for the sorting function. The test utilizes a separate dataset from the one used in the CAGRA main test to avoid the effect of rounding errors during norm computation between two vectors in the dataset. More details are in the source code.
https://github.com/enp1s0/raft/blob/ea6c449c260895e9125a591a4848eed06f5b72c4/cpp/test/neighbors/ann_cagra.cuh#L93-L96

Issue

Close #1446

[enp1s0]

@tfeher Can you check the PR and let me know if there are any problems?

[tfeher]

Thanks @enp1s0 for the PR! It looks good overall, below you will find a few small comments (and just as a reference, I have also linked two existing issues).

[enp1s0]

@tfeher Thank you for your comments.
I have added a code example of sort_knn_graph and fixed the stream synchronization in the test code. I'll fix the memory allocation and multi-GPU support in another PR because there will be a lot of code changes.

[tfeher]

Thanks @enp1s0 for addressing the issues. The PR looks good to me.

[enp1s0]

I have changed the test case parameters to satisfy the following requirements in int8 and uint8 tests.

https://github.com/rapidsai/raft/blob/20141120f6ae2fd9ea1ea0648507af04594921ec/cpp/include/raft/neighbors/detail/cagra/cagra_build.cuh#LL52C45-L52C45

RAFT_EXPECTS(
    dataset.extent(1) * sizeof(DataT) % 8 == 0,
    "Dataset rows are expected to have at least 8 bytes alignment. Try padding feature dims.");

[cjnolet]

/merge