deepseek41 : report n_swa = 0 like deepseek4 so the server keeps its checkpoints

[midagedev]

llama_model_n_swa() returns 0 for LLM_ARCH_DEEPSEEK4 because the dsv4 cache's 128-token raw window cannot be used for rollback (the comment in the function says why). LLM_ARCH_DEEPSEEK41 was added next to it without extending that case — every other DEEPSEEK4 special case in src/ (graph selection, KV cache construction, the is_dsv4 checks in llama-arch.cpp and llama-context.cpp) already lists both. This one line does the same here.

Effect without it. llama-server computes its checkpoint threshold as pos_next - n_swa, with n_swa = 128 for V4.1, and rejects every context checkpoint within 128 tokens of the resume point — including the one it creates 4 tokens before the end of the prompt. Each follow-up turn on a long conversation resumes from the previous checkpoint, about 2 048 tokens earlier, and re-prefills that stretch.

Measured on DeepSeek-V4.1-Flash Q3_K_M (12k-token document, then four ~550-token follow-ups, temperature 0, cache_prompt on; RTX A6000 + RTX 3090 + 5975WX, -c 262144 -ub 2048 -ctk q8_0 -ctv q8_0, DSpark draft):

	turns 2–5 prompt_n	turns 2–5 TTFT	answers
runtime/deepseek41 at 5210c7c	2606 / 2587 / 2618 / 2579	20.6 / 21.1 / 21.3 / 21.1 s	—
+ this commit	562 / 543 / 574 / 535	8.9 / 8.8 / 8.8 / 8.2 s	5/5 byte-identical

Trace before (-lv 5, turn 2):

checking checkpoint with [11939, 11939] against 11816   <- 11944 - 128, rejected
checking checkpoint with [9895, 9895] against 11816     <- restored

after:

checking checkpoint with [11939, 11939] against 11944   <- restored

Mainline master with DeepSeek-V4-Flash restores the same 4-tokens-before-the-end checkpoint on every turn (measured, 862 tokens per follow-up), so this brings V4.1 to the behaviour V4 already has.

The same symptom can be fixed on the server side by treating pos_min == pos_max checkpoints as exact states (which is what ggml-org#25592 does for hybrid/recurrent memory); I tried that first and it measures the same, but for this model the model-side special case already exists and is the intended mechanism, so this is the smaller change.

If this branch gets squashed before it goes upstream, I'd appreciate the commit (or a Co-authored-by) surviving — no worries if the history has to be rewritten.

Prepared with AI assistance (Claude); the measurements were run by me.

[midagedev]

Carried into ggml-org#28696 as 3b1179f. Thanks for picking it up.