Apply crawl_delay per request instead of per page

[thatbudakguy]

Closes #209

Moves the requests a harvest makes into a GeoBlacklightHarvester::HttpClient that waits out the crawl delay before each request. Previously the delay was applied once per page of search results, so a harvest could issue 100 document requests back to back before it took effect. This resulted in ~20% of requests failing outright against our local firewall at Stanford, for example.

Each request continues to get its own connection rather than reusing one, which is an architectural difference from something like GitHub downloading, where we instead want a persistent connection pool (using Faraday). Firewalls are somewhat less likely to reject a request if it's a new request every time instead of a persistent connection making big bursts of requests.

Fractional delays work too, so you can tune things to get an acceptable timing for large harvests.

Harvest timing changes

Pacing per request is the point of #209, but it makes a harvest considerably slower than it was: a 1 second delay against a GeoBlacklight with 10,000 records is roughly three hours of waiting, where before it was ~100 page delays.

GeoCombine is much more useful when doing incremental harvesting instead of full dumps anyway, so the assumption is that longer delays are acceptable when you're pulling directly from a GeoBlacklight (especially if it's your own). The "happy path" is getting it from the OGM GitHub anyway, which is much faster.