Consolidate SUM reductions

[mfoerste4]

This PR consists of multiple parts:

1. redirect custom reduction kernels within stats namespace to linalg::reduce
2. Specialize reduction kernels for addition utilizing the Kahan-Babushka-Neumaier-Sum link
3. Slightly adjust kernel heuristics for coalesced reductions

This should address #2366 and #2205. With the kernel heuristics adjusted the maximum performance drop is 4%.

FYI, @tfeher

[tfeher]

Thanks Malte, this looks really good! I love how a new compensated reduction is reused among various stat functions! I just have a small request to improve documentation.

[mfoerste4]

Thanks Malte, this looks really good! I love how a new compensated reduction is reused among various stat functions! I just have a small request to improve documentation.

Thanks @tfeher for reviewing. I added a note to both the implementations in detail as well as the main entry in raft::linalg::reduce.

[tfeher]

Thanks @mfoerste4 for the update, LGTM!

[mfoerste4]

@tfeher , sorry for the late change, I missed this before. The sample-parameter for mean/stdev was not implemented for column based data previously. I enabled it and updated the failing tests (that were sampling with rows == 1)

[tfeher]

Approving latest changes.

[mfoerste4]

This raised a bug in the test code that existed for a long time. IIUC the covariance algorithm using the pre-computed mean should always use the standard mean, no matter whether we are computing the sampled covariance or not.
A quick scan of the cuml brought up this code which might be wrong as well.

[tfeher]

Thanks for catching these issues, LGTM.

[cjnolet]

/merge