[REVIEW] Fix memory leak and unnecessary allocs in TSNE::get_distances

[zbjornson]

c3ab759 left behind two extra allocations. Those allocations also leaked because delete knn_input, sizes only deletes knn_input. (It should also be delete[].)

Finally, modernizes the vector initialization syntax.

[GPUtester]

Can one of the admins verify this patch?

[cjnolet]

We could add a convenience wrapper for the brute force kNN that accepts a single float* and calls the version that accepts multiple. Would you like to do that in this PR?

[zbjornson]

Oops, unclear comment. I was only trying to avoid the const_casts here and on line 58R, not the vector. But it looks like making it const-correct would be hard, so I've removed the comment.

[cjnolet]

Changes LGTM. Thanks for the contribution!

[JohnZed]

Ok to test

[zbjornson]

Just realized that this fixed another bug that caused double the memory usage.

std::vector<float *> input_vec(1); constructs a vector with 1 float default-inserted (definition 4), then push_back() expands the vector so it has size 2. Downstream, the knn code branches if input.size() > 1 to allocate a lot of temp memory (and do extra work):

cuml/cpp/src_prims/selection/knn.cuh

Lines 263 to 265 in 069a229

	if (input.size() > 1) {
	all_D.resize(input.size() * k * n, userStream);
	all_I.resize(input.size() * k * n, userStream);

The only other place I see that issue is here, but it's in a test, so not necessarily wrong:

cuml/cpp/test/prims/knn.cu

Lines 70 to 73 in 069a229

	std::vector<float *> input_vec(1);
	std::vector<int> sizes_vec(1);
	input_vec.push_back(ptrs[0]);
	sizes_vec.push_back(sizes[0]);

[cjnolet]

@zbjornson feel free to make the change to the kNN test in #2548.