Avoid dependency inversion on pool exhausted

[carrascoacd]

Problem

When using Finch is expected to receive a RuntimeError if the connection pool is exhausted, see https://github.com/sneako/finch/blob/main/test/finch_test.exs#L561

This is a problem because in a service that uses a lower tier API than the service itself, a dependency inversion of tiers can happen. For example, a tier 2 service calling a tier 3 service.

The tier inversion happens because we are not rescuing the error and returning it into a tuple, so the operation can work without problems under this situation.

Proposed solution

For clients using Tesla: elixir-tesla/tesla#744

Unresolved questions

• Is this behavior expected, and why?
• Do you think this reraise should be changed to return an error tuple instead, so we don't provoke the linked caller to exit?

cc/ @yordis @sneako