Simplify delta-net

[ikawrakow]

With PR #1333 merged, and the fused delta-net implementation strictly better than autoregressive or chunked, there is no need to keep these two around. This PR removes them, and also removes the -fdn | --fused-delta-net command line argument.

[ubergarm]

Did a quick sweep-bench on Qwen3.5-27B Dense and running well with -sm graph being faster than -sm layer.

👈 Details

-sm graph

CUDA_VISIBLE_DEVICES="0,1" \
./build/bin/llama-sweep-bench \
  -n 128 \
  --warmup-batch \
  --model "$model" \
  -c 69632 \
  -ger \
  -sm graph \
  -ngl 99 \
  -ub 4096 -b 4096 \
  --threads 1 \
--no-mmap

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
4096	128	0	2.582	1586.35	3.310	38.67
4096	128	4096	2.631	1556.58	3.252	39.37
4096	128	8192	2.678	1529.51	3.288	38.93
4096	128	12288	2.728	1501.24	3.322	38.53
4096	128	16384	2.794	1465.98	3.381	37.86
4096	128	20480	2.846	1439.10	3.402	37.62
4096	128	24576	2.919	1403.19	3.425	37.38
4096	128	28672	2.983	1373.13	3.448	37.13
4096	128	32768	3.031	1351.58	3.497	36.61
4096	128	36864	3.089	1326.05	3.517	36.39
4096	128	40960	3.144	1302.94	3.540	36.16
4096	128	45056	3.208	1276.80	3.560	35.95
4096	128	49152	3.270	1252.78	3.607	35.49
4096	128	53248	3.320	1233.66	3.628	35.28
4096	128	57344	3.376	1213.32	3.651	35.05
4096	128	61440	3.435	1192.31	3.670	34.88
4096	128	65536	3.501	1169.89	3.717	34.44

-sm layer

CUDA_VISIBLE_DEVICES="0,1" \
./build/bin/llama-sweep-bench \
  -n 128 \
  --warmup-batch \
  --model "$model" \
  -c 69632 \
  -ger \
  --merge-qkv \
  -sm layer \
  -ngl 99 \
  -ub 4096 -b 4096 \
  --threads 1 \
  --no-mmap

PP	TG	N_KV	T_PP s	S_PP t/s	T_TG s	S_TG t/s
4096	128	0	2.970	1379.08	3.551	36.05
4096	128	4096	3.071	1333.65	3.533	36.23
4096	128	8192	3.161	1295.95	3.612	35.44
4096	128	12288	3.278	1249.61	3.646	35.11
4096	128	16384	3.388	1209.10	3.713	34.47
4096	128	20480	3.496	1171.73	3.752	34.11
4096	128	24576	3.608	1135.25	3.821	33.50
4096	128	28672	3.717	1102.02	3.858	33.18
4096	128	32768	3.822	1071.55	3.925	32.61
4096	128	36864	3.927	1042.93	3.959	32.33
4096	128	40960	4.030	1016.27	4.025	31.80
4096	128	45056	4.136	990.22	4.064	31.50
4096	128	49152	4.240	966.08	4.128	31.01
4096	128	53248	4.346	942.47	4.164	30.74
4096	128	57344	4.463	917.86	4.231	30.25
4096	128	61440	4.551	900.07	4.265	30.01
4096	128	65536	4.650	880.88	4.330	29.56