RTDS snapshot should be pinned to each stream

*Title*: we should keep runtime's consistency during a request's lifecycle.

*Description*:
runtime layers should be consistent during processing of a request.

*Repro steps*:
> a bug scenario:
filter A reads version 1 of runtime snapshot, filter A stopIteration, waits for a some callback result, say on-demand-rds
a new Runtime snapshot (version 2) is pushed by RTDS,
filter A resume the filter iteration, and filter B now sees version 2 when it reads from loader.runtime().snapshot()

IMO, We should latch the runtime snapshot to a stream during its creation time.