Denying a tool during a 'permission stop' now cleanly ends the turn

What

• A new abort/interrupt reason, permission-stop, was added alongside the existing interrupt, turn-abort, and refusal-fallback-edit reasons, with its own error mapping.
• When a tool call is denied while the abort reason is permission-stop, the resulting tool-result message now carries a toolDenialEndsTurn flag. A new helper walks backward through the conversation and treats a user message with this flag set as one that ended the turn, rather than one that was merely interrupted.
• This toolDenialEndsTurn flag is now carried through the rest of the tool-result plumbing (alongside existing fields like toolDenialKind/userFeedback) and is checked to short-circuit processing when a denial should end the turn.
• Subagents are affected too: when a synchronous subagent is stopped for this reason, it's now recorded distinctly (permission_stop_sync) in telemetry and cleaned up as a system action rather than being reported as a plain user cancellation.

Why

Previously, denying a tool call during a permission stop could be misclassified as a generic interruption. This change makes Claude Code correctly recognize that the turn ended because of the permission denial, so the conversation transcript and telemetry reflect what actually happened.