llama-memory: enable checkpointing with partial rollback

[am17an]

No description provided.

[ggerganov]

cell.src >= 0 ? cell.src

I guess this was the part that I was missing in my approach. I will take detailed look tomorrow and run some tests to verify it's working as expected.

[zzhenyao]

Tested on Qwen3.6-27B. Checkpoint miss from ggml-org#22673 (in master and mtp-clean) is fixed here. nice work!.👍

11.26.676.492 I slot launch_slot_: id  0 | task 4631 | processing task, is_child = 0
11.26.676.549 W slot update_slots: id  0 | task 4631 | n_past = 101412, slot.prompt.tokens.size() = 101743, seq_id = 0, pos_min = 101742, n_swa = 0
11.26.676.549 I slot update_slots: id  0 | task 4631 | Checking checkpoint with [101409, 101409] against 101412...
11.26.731.828 W slot update_slots: id  0 | task 4631 | restored context checkpoint (pos_min = 101409, pos_max = 101409, n_tokens = 101410, n_past = 101410, size = 262.973 MiB)
11.27.422.622 W slot create_check: id  0 | task 4631 | erasing old context checkpoint (pos_min = 94538, pos_max = 94538, n_tokens = 94539, size = 255.293 MiB)
11.27.533.495 I slot create_check: id  0 | task 4631 | created context checkpoint 24 of 24 (pos_min = 101752, pos_max = 101752, n_tokens = 101753, size = 263.356 MiB)