qwen35 : in-place recurrent state rows for plain decode (GGML_GDN_ROWS_PLAIN=1)

[jasontitus]

Overview

The fused gated delta net has a rows mode that reads and writes each sequence's state rows in the recurrent cache in place. Today it is used only when n_rs_seq > 0 (speculative contexts with snapshots). Plain decode gathers every layer's state, runs the recurrence, and copies the state back; on the Bonsai 2 models that is 10-14% of a decode token on an M5 Max.

With GGML_GDN_ROWS_PLAIN=1, plain decode takes the rows path too, with one snapshot slot (K = 1). Output is bitwise identical. Off by default.

• Only when no extra cells are relocated (n_rs == n_seqs): the relocation in build_rs_cache_view runs before the GDN read and, after a cell reorder, could overwrite a row another sequence reads (the gathered path reads first). Graph reuse compares the s_copy_extra size, so such a batch rebuilds and takes the gathered path.
• GGML_GDN_ROWS_PLAIN_MAX_TOKENS=N limits it to at most N tokens per sequence: on an A19 the in-place recurrence op is ~18% slower at 512-token prefill while decode gains. Contexts with n_rs_seq > 0 keep rows mode at every width, as today.
• delta-net-base.cpp accepts rows mode without snapshots (the assert that required n_rs_seq > 0 becomes the keep condition).

Related: #207 does an in-place GDN update for the CPU backend with a new op. This PR only reuses the existing rows path (Metal) for plain decode; the two touch the same lines in qwen35.cpp and delta-net-base.cpp.

Additional information

M5 Max, Ternary Bonsai 2 27B PTQ1_0, this branch, switch off vs on, three A-B-B-A quartets with 8 s cooldowns (tok/s; llama-server rates are aggregate over the requests, 128 greedy tokens):

Case	Off	On	Speedup [quartet range]
tg128 (llama-bench)	38.95	41.96	1.076 [1.068-1.091]
pp512 (llama-bench)	722.8	755.9	1.046 [1.042-1.052]
llama-server, 1 request	37.49	40.13	1.070 [1.065-1.073]
llama-server, 2 requests	18.64	19.25	1.033 [1.021-1.040]

Generated tokens identical in every pair. Every logit bitwise equal off vs on for 96 single-token decodes plus 12 batches of 4 on PTQ1_0, Q1_0 and PQ2_0 Bonsai models.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. Claude (Anthropic) helped with the implementation, testing, measurements and this description; I reviewed the change and own it.