fix(storage): make TTL deletion error propagation opt-in

[cocolord]

Description

Follow-up to merged PR #5268, which made Volcengine AK/SK scalar-query and delete errors propagate for all callers. Restore the original default behavior and add internal VikingFS.rm(..., raise_on_error=True) for TTL callers that need to observe those request failures. The option defaults to false and is passed explicitly through the existing URI deletion, adapter and collection calls; no public HTTP, SDK, CLI or MCP option is added.

The opt-in covers both ID enumeration and delete requests, reusing the existing HTTP/JSON error handling. Other backends retain their contracts. An accepted deletion request is sufficient: no synchronous index count, residue checks, extra locks or stored fields are added. TTL failure registration and retry scheduling remain outside this change.

Validation

• One parameterized regression in the existing native filesystem test file exercises the full VikingFS.rm → vector backend → adapter → Volcengine request path. Its 16 combinations cover default/opt-in behavior, scalar-query/delete failures (HTTP 429, HTTP 500, invalid JSON), accepted requests, repeated deletion of absent files, and retaining files on opt-in failure. Ordinary scalar queries and delete-all retain their defaults after opt-in calls.
• The earlier adapter-only regression is replaced by this end-to-end regression; no new test file is added.
• On current main (ec56b9443), uv run --no-sync pytest -q -o addopts= tests/storage/test_volcengine_clients.py tests/misc/test_vikingfs_uri_guard.py tests/storage/test_transfer_merge_binding.py -k 'volcengine or http or uri or transfer_uses_queries_without_count_or_sort or rm_cloud_errors_are_opt_in': 106 passed, including all 16 new combinations and 90 existing related cases.
• Changed Python files pass Ruff lint/format and git diff --check. Targeted mypy with --follow-imports=silent reports the same 234 existing diagnostics as the merge-base snapshot, with no added or removed diagnostics; it is not green. The added regression has no type errors. Validation used the existing uv environment.
• No live remote-backend deletion was performed.

Human Involvement

• A human participated in the implementation or review loop
• This PR was generated entirely by AI agents without human participation in the loop