Fixes/enhancements

[alyst]

Another bunch of fixes and enhancements

• scale the input matrix to avoid degeneration during perplexity calculation
• Hbeta(): handle situations when the point is too distant from any other points; previously it resulted in a degenerated P distribution
• further reduce the number of array re-allocations
• rename error to KL-divergence, optimize the code, output the divergence as a part of progress meter
• correct the KL-divergence for the first few iterations when we are cheating with the P matrix scale
• add pca_init=true/false kwarg to tsne(). If true, the initial layout is the first ndims components of pca(X). It should be a better start than a random layout (in terms of KL-divergence). It also makes the tSNE layouts reproducible.

[alyst]

@lejon Thanks, finally it passed the tests. Yet I don't know why it takes so long for 0.4 and 0.5 builds. Locally I have 0.5 and it's running fine.

BTW, now that the package is updated to Julia 0.5, it would be great to proceed with #3. I think many people would appreciate that.