Investigate if we should use near-fetch

[dndll]

We are reimplementing a bunch of stuff near-fetch already has - Ideally we just contribute work there (unless they're using incompatible runtimes), but we should try to reach a sovereign client that isn't broken by these things.

near-fetch also already has:

• memoization of nonces
• retryability

If it doesn't have archival failover, we should probably roll up our stuff into that.