TODO

* apply golint and errorcheck

* right now every reader is expected to instantiate a Store data-structure
  which will open a new instance of the index/kvfile for reading.

* when number of nodes in leaf-cache exceeds `MaxLeafCache`, figure out a
  cache eviction algorithm is efficient for btree.

* I am not sure locking caches for intermediate nodes and leaf nodes for
  readers will have impact on scalability (especially in cases of
  large number of cores).

* at present count of all entries under a sub-tree or even under the root tree
* had to be computed by walking down all the leaf nodes under the sub-tree.
  can we include another field called `count` inside the intermediate node's
  structure ?

* define Range() API.

* right now, remove works only for individual entries identified by
  {key,docid} tuple, modify its specification to remove all entries
  identiefied by {key}.

* lookup() and other traversal apis that use channel to return back the
  result to caller can used buffered-channel to avoid blocking on mvQ.

* Long outstanding reads can block snapshots being flushed and stales nodes
  reclaimed.

* Optimize flushSnapshot() to flush only leaf nodes. And periodically flush
  the entire cache for intermediate nodes. This could mean that the in-memory
  snapshot reference for reads will be ahead of on-disk snapshot reference.

Coding guidelines

1. Using _ assignment to avoid unsided imports.

2. Avoid using fmt.* in test cases, instead use t.Error(), t.Errorf() etc ...

3. Don't name test functions with `_`, use camelCase.

4. panic(err) is sufficient, don't need to do panic(err.Error())