fix: reject stale conversation turns

[franciscojavierarceo]

Summary

• capture a typed conversation version alongside rehydrated history
• reject stale explicit-conversation turns under the existing row lock before persistence
• return a consistent conversation_locked error over HTTP, SSE, and WebSocket, including generate: false
• preserve previous_response_id fork behavior and the existing unchecked storage API
• add SQLite, transport, and two-pool PostgreSQL race coverage plus deployment documentation

This builds on the PostgreSQL persistence foundation merged in #142.

This implements optimistic conflict detection at persistence time. A pre-inference distributed lease or queue that also suppresses external tool side effects remains separate policy work.

Test Plan

• cargo fmt --all -- --check
• cargo clippy --all-targets -- -D warnings
• cargo test
• pre-commit run --all-files
• postgres_concurrent_conversation_writes_have_contiguous_sequences against PostgreSQL 17
• postgres_optimistic_conversation_conflict against PostgreSQL 17
• postgres_lock_wait_is_bounded_without_blocking_other_conversations against PostgreSQL 17
• read-only Claude review: no actionable findings
• gstack pre-landing review: all in-scope actionable findings resolved

[maralbahari]

@franciscojavierarceo can retarget to main?

[maralbahari]

converts StorageError::ConversationConflict { conversation_id } into the unit-like ExecutorError::ConversationLocked. Consequently, persist.rs (line 30) logs only a generic message, and Error::source() cannot reach the original typed error. This makes concurrent-write incidents impossible to correlate with a conversation. Preserve StorageError as a #[source] while keeping the client-safe response.