feat: manage workspace inventory, deletion, and stale cleanup

[harshitethic]

Problem

PatchPilot creates a new on-disk clone for every agent run, but there is currently no lifecycle API for those workspaces. Long-running local instances can therefore accumulate clones indefinitely, and clients cannot enumerate or explicitly dispose of completed runs.

What changed

Adds a contained workspace lifecycle layer:

• GET /api/workspaces — list valid workspaces with current branch, dirty-file count, and UTC update time;
• DELETE /api/workspaces/{workspace_id} — remove one workspace;
• POST /api/workspaces/cleanup — remove workspaces older than a bounded age (1 hour to 30 days).

Safety

Workspace directory resolution is centralized in workspace_dir() and constrained beneath .workspaces before any read or delete operation.

Cleanup independently resolves every candidate beneath the workspace root before deletion, skips invalid/unreadable entries, and only removes entries older than the requested cutoff.

Tests

Regression coverage verifies:

• cleanup removes expired workspaces while preserving fresh ones;
• explicit deletion removes only the selected workspace;
• traversal attempts such as ../outside are rejected and leave external directories untouched;
• invalid cleanup windows are rejected by Pydantic.

The existing single-workspace inspection API remains unchanged. Documentation includes the new lifecycle endpoints and retention semantics.