Support for cosine similarity search on sparse vectors

[vamshin]

By default k-NN plugin uses dense vector for Cosine similarity measure. NMS library has support for sparse vector space type. We could enable this support in k-NN plugin.

[jmazanec15]

In order to implement Sparse Vector Support, we would need to make a modification to our object creation in JNI code. Currently, the dataset is represented as float * and that pointer is passed in to create the corresponding object. However, for sparse vector support, we would need to represent the Object's data set as SparseVectElem.

Additionally, we need to think of the appropriate UX for users to input data for this format. For example, users could input data in the following format:
{<index>: <value>,..}

[chenqi0805]

I can take on this issue if it is not blocking. Together with #109

Additionally, we need to think of the appropriate UX for users to input data for this format.

@vamshin @jmazanec15 Shall we come up with a design of an appropriate sparse vector input format or there is UX designer doing the job?

[chenqi0805]

BTW, in nmslib, sparse space type does not support optmized index, either(#161 )

[chenqi0805]

IMO, we should add a new field type knn_sparse_vector. This is worth a design doc before execution. also applies to #109