fix: Honor Retry-After on 5xx responses, not just 429 (#606)

[ffumero2003]

Problem

A 429 with Retry-After: 2 makes the client wait ~2s, but a 503 (or any 5xx)
with the same header waits only its default backoff — the header is ignored.

Root cause

interpret_b2_error's 5xx branch builds ServiceError('%d %s %s'...) from the
message only and never reads response_headers['retry-after']. ServiceError
had no __init__, so it inherited B2Error's unconditional
retry_after_seconds = None. b2http's retry loop then always falls back to
default backoff for 5xx. (429 works because TooManyRequests stores the header.)

Fix

Mirror TooManyRequests: give ServiceError an __init__ that accepts and
stores retry_after_seconds, and pass response_headers.get('retry-after')
through in the 5xx branch. The read side (b2http._translate_and_retry) already
honors a non-None value, so no change there.

Testing

• New unit tests: 503 with a Retry-After header stores it; without one it stays
  None (red before, green after), mirroring the existing 429 tests.
• Full unit suite green across all API versions (nox -s unit, apiver v0–v3).

Fixes #606.

[sophiecarreras]

This is superseded by #619, which carries the same code and tests (only the changelog sentence differs). Closing this one avoids two PRs adding the same fragment and test. Please close whichever you want to drop; I suggest keeping #619.

[ffumero2003]

Superseded by #619, which merged to master with the same fix plus the review nits. The change here is already on master, so closing this duplicate.