fix: persistent disk cleanup and fallback behavior (⚠️ UNTESTED)

[wdvr]

⚠️ WARNING: UNTESTED CHANGES

These changes fix real bugs but have NOT been tested in production. Review carefully and test before merging.

---

Fix 1: Force-detach EBS volumes before deletion

File: lambda/reservation_expiry/index.py (+15 lines)

Problem: When cleaning up expired reservations with persistent disks, volume deletion fails with VolumeInUse error if the volume is still attached to an instance.

Solution:

• Force-detach the volume before deletion
• Wait for volume to reach "available" state (up to 2 minutes)
• Then delete the volume

Impact: Prevents orphaned volumes that accumulate costs when cleanup fails.

---

Fix 2: Don't silently fall back to EmptyDir when user requests persistent disk

File: lambda/reservation_processor/index.py (±8 lines)

Problem: When a user explicitly requests a persistent disk via disk_name parameter, if that disk setup fails (e.g., disk doesn't exist, disk in use), the code would silently fall back to EmptyDir instead of failing the reservation.

Solution:

• If disk_name is set by user → fail the reservation with clear error message
• If error is "disk in use" → fail the reservation (regardless of disk_name)
• Only fall back to EmptyDir for old-style reservations without explicit disk_name (backward compatibility)

Impact: Users get clear error messages when their disk request fails, instead of silently getting EmptyDir when they expected persistent storage.

---

Test Plan (REQUIRED before merge)

• Test volume cleanup: Create reservation with persistent disk, let it expire, verify volume is detached and deleted successfully
• Test snapshot flow: Create reservation with persistent disk, let it expire with snapshot requested, verify snapshot completes and volume is cleaned up
• Test explicit disk_name: Request reservation with disk_name for non-existent disk, verify reservation fails with clear error (not EmptyDir fallback)
• Test disk in use: Request reservation with disk that's already attached, verify reservation fails with clear error
• Test backward compatibility: Old-style reservation without disk_name parameter should still fall back to EmptyDir on disk errors
• Monitor for orphaned volumes: Check AWS console for any EBS volumes not cleaned up properly

---

Files Changed

• terraform-gpu-devservers/lambda/reservation_expiry/index.py: Add detach-before-delete logic
• terraform-gpu-devservers/lambda/reservation_processor/index.py: Check disk_name before fallback

Do NOT merge until tested!