Fix an issue with PCA not working with Float32 arrays

[maximsch2]

Call to fit(PCA, X) fails if X is Float32 array. This pull request adds a test and fixes it. I'm also fixing pcacov and pcasvd as they seem to have the same issue.

[maximsch2]

Those build failures seem to be due to issues in JSON.jl and not the PR.

[alyst]

pr should still retain FloatingPoint type declaration (here and in the functions below)

[maximsch2]

Thanks for the comments! I've added more tests, removed convert and put FloatingPoint there.

[alyst]

@maximsch2 Thanks. Could you please also update the declaration of pratio in `pcacov()and then squash your commits into a single one? The bigger remaining issue is that now the nightly Julia branch fails inpcacov()`` in one of your newly added tests. Have you looked into that?

[maximsch2]

Fixed & squashed. I've tried reproducing the error that travis gets, but I don't get it here.

[maximsch2]

Added a fix for deprecated FloatingPoint in pca.jl as well. Tests pass locally for me, on 0.3 on Travis and 0.5-dev on Travis seems to be broken in an unrelated way.
@alyst Can you take a look at it?

[jiahao]

I believe the test failure on the nightlies was caused by JuliaLang/julia#13465

/cc @andreasnoack please have a look

[andreasnoack]

Wow. That is a lot of warnings, but I found the error, and it is indeed because of JuliaLang/julia#13465. I'll have to take another look at the deprecation logic.

[andreasnoack]

I believe JuliaLang/julia#13505 should fix this issue

[andreasnoack]

@maximsch2 Would you try to rebase your branch on top of master and push force to your fork. I've made some changes to the .travis.yml file that might fix the test here. Thanks.

[maximsch2]

Hm, now the test fails on 0.4 on Linux (it works on OS X on travis as well as on Ubuntu 14.04 locally here). It looks like that's probably just a rounding error. Do we really need that check?

[andreasnoack]

Maybe the check will catch a blunder some day, but I don't this it is that important. You could also fix the test error by limiting the number of principal components to d-1 or by setting tval=max(vsum,sum(v)). I think you can just choose the one you prefer.

[maximsch2]

Added maxoutdim=3 to the test (nearby tests seem to do the same). Now Travis is happy.

[andreasnoack]

Thanks.

[maximsch2]

Awesome, thanks!