fix: retry transient get_entities failures during observation

[JackHopkins]

Problem

During a 4-node, 64-step eval, one rollout died ~1h in with "Error getting observation: Error getting entities while getting observation" and --fail-on-error cancelled the other three healthy epochs. Two compounding issues:

1. The wrapper in FactorioGymEnv.get_observation() discarded the underlying exception message, so the eval log contains no diagnosable cause.
2. A single transient failure (the full 41-program action sequence replayed clean on a fresh server, twice) is fatal to the whole epoch — and with fail-on-error, to the whole run.

Fix

• Retry get_entities up to 3 times (0s/2s/5s backoff) before failing the observation, logging each attempt.
• Include the underlying error text in the raised exception so real failures are diagnosable from the log.

Verification

• Lint/format pass with CI's ruff; environment imports clean; tests/gym_env/test_server_reuse.py passes (3/3)
• Failure mode reproduced-by-absence: replaying both failed epochs' exact program sequences (47 and 41 programs) against live servers produced zero get_entities failures, confirming transience