Prevent unsafe retries for ambiguous writes

[SahirVhora]

What changed

• retry transient failures automatically only for safe HTTP methods
• raise AmbiguousWriteError when a write receives a retryable server response or loses its response to a network error
• export the new exception for downstream tools
• add regression coverage proving POST requests are attempted exactly once

Why

The shared retry loop previously replayed POST, PATCH, and DELETE requests. If SuccessFactors committed a mutation but its response was lost, replaying the request could duplicate the change or turn a successful create into a misleading duplicate failure.

Downstream callers can now reconcile target state before deciding whether a write should be retried.

Validation

• pytest: 175 passed, 2 skipped
• scoped Ruff checks: passed
• git diff --check: passed

The repository-wide mypy run still reports existing issues in unrelated modules and missing third-party stubs.