Request deduplication

[dchoi-viant]

With the implied batch expansion, there are some requests that can be deduplicated.
Ideally, this would be done on the client side, but the server side can do it as well.

[dchoi-viant]

After inspection in #37, one aspect to consider is how CacheStatusFoundNoSuchKey operates.

While not a thorough solution for request deduplication (as it only applies to non-batch and cacheable requests), it can reduce redundant model predictions. See https://github.com/viant/mly/blob/v0.16.x/WORKFLOW.md#run-full-cache-miss - instead of having CacheStatusFoundNoSuchKey from the client-side re-submit a prediction request, it should instead use singleflight to wait for the initial request sent to complete.
This should be safe since CacheStatusFoundNoSuchKey only should exist in the case there is a local scache miss, meaning another Goroutine must have already send a prediction request to the server.