[BUG] Barnes-Hut tSNE does not obey initialize_embeddings param

[zbjornson]

Describe the bug

TSNE_fit (and its Python interface) takes a parameter initialize_embeddings:

cuml/cpp/include/cuml/manifold/tsne.h

Lines 62 to 63 in 069a229

	* @param[in] intialize_embeddings Whether to overwrite the current Y vector
	* with random noise.

That's obeyed by TSNE::Exact_TSNE but not by TSNE::Barnes_Hut, which doesn't even have that as a function parameter.

Compare to exact's impl:

cuml/cpp/src/tsne/exact_tsne.cuh

Lines 61 to 62 in 069a229

	if (intialize_embeddings)
	random_vector(Y, -0.0001f, 0.0001f, n * dim, stream, random_state);

(Realize as I write this up that it's also misspelled.)

Steps/Code to reproduce bug
(Found by code review, don't have a repro.)

Expected behavior
Y should not be overwritten.

Environment details (please complete the following information):

• Environment location: all
• Linux Distro/Architecture: all
• GPU Model/Driver: all
• CUDA: all
• Method of cuDF & cuML install: from source, build-independent

Additional context
This is easy to fix. I can do it once my other tSNE PRs are merged.

[drobison00]

@zbjornson Are you planning on submitting this with/after #2620 ?

[zbjornson]

@drobison00 the above commit (6581922) fixes this issue, but I want to test it more before opening a PR.

#2620 requires a more substantial change since B-H doesn't calculate the gradient at all right now, AFAIK. I don't have immediate plans to work on it.

I do want to work on K-L divergence (#863). ("want" as in I have a pressing need for that feature, but not a desire per se to implement it. I can help review a PR if someone else works on it before I get to it.)