Instead of using in-process caching for tokens per service per
worker (disabled by default now), use a shared memcache to cache
token validation(s). This should both offload/speedup validations
and avoid the issues surrounding inconsistent validation responses
when using in-process caching [since each worker caches separately].
Change-Id: Ifc17c27744dac5ad55e84752ca6f68169c2f5a86