Support for sparse feature indices

[Sally](http://www.mlsec.org/sally/) generates LibSVM vector spaces like

```
1 1191618:0.973645 1243483:0.362496 1346494:0.995527
0 1243483:0.226674 1346494:0.997515 2362908:1 4604526:1
```

where feature indices are really (murmur3?) hashes of words.

Maybe we could do

```
Svm::Model#predict(1191618 => 0.973645, 1243483 => 0.362496 [etc]
```
