implement async kvs_fence() call

[garlick]

As noted in #706, we need a version of kvs_fence() that is non-blocking, and allows the response to be handled via an RPC continuation.

The KVS API is due for a rework anyway. I propose we go ahead and take a step in this direction by reworking kvs_fence(). Where I've been stuck before is in naming the two halves, especially when we'll have to split kvs_get(), since the second half might naturally have a _get suffix to match flux_rpc_get(). Her's a tentaive proposal:

Current blocking prototype:

int kvs_fence (flux_t h, const char *name, int nprocs);

Proposed new prototypes:

// calls flux_rpc()
flux_rpc_t *flux_kvs_fence_start (flux_t h, const char *name, int nprocs);

// calls flux_rpc_get()
int flux_kvs_fence_finish (flux_rpc_t *rpc);

// calls flux_kvs_fence_start() then flux_kvs_fence_finish()
int flux_kvs_fence (flux_t h, const char *name, int nprocs);