support global storage as cache locations

[adammoody]

While it should be possible to use global storage as a cache level, we currently have some hard-coded assumptions and scalability issues because SCR generally assumes cache locations are using node-local storage.

For example, when inspecting the cache on restart, one rank is responsible for distributing work. With global storage, this means rank 0 will have to talk with every process in the job, which becomes an N-to-1 bottleneck.

We should extend SCR to efficiently support global cache locations.

Related to #133

[adammoody]

Replaced with more specific issues:
#489
#493
#494

[mShining]

这是来自QQ邮箱的假期自动回复邮件。   您好，我最近正在休假中，无法亲自回复您的邮件。我将在假期结束后，尽快给您回复。