Built-in retry/backoff for transient gateway 5xx in API.send_request

[martinkersner]

Problem

tests/conftest.py wraps API.send_request to retry 502/503/504 because the prod edge "intermittently returns these under load — infra flakiness, not an SDK bug." That retry exists only in the test suite, so real consumers hit the same transient 5xx and get a ServerError on the first blip, with no handling.

Proposal

Move bounded retry-with-backoff into datamaxi/api.py::API.send_request (or _dispatch_request):

• Retry a small, configurable set of transient statuses (default 502, 503, 504) and connection/read timeouts.
• Linear or exponential backoff, capped attempts (e.g. 3).
• Opt-out / tunable via constructor kwargs (e.g. max_retries=, retry_statuses=).
• Non-transient 4xx/5xx still raise immediately.

Notes

• Once in the SDK, the conftest fixture can be simplified/removed.
• Keep it dependency-free (loop + time.sleep) or consider urllib3.util.retry.Retry on the shared Session (one Session now exists after refactor: share one Session across the client tree via composition (#2, #3) #137).