CUDA: cache intermediate tensors

[am17an]

Cache quantized activations and mul_mat_id_helper output for the gate tensor. Modest TG and PP gains

[JohannesGaessler]

I think the bigger potential to do with quantizing the activations lies in doing the quantization prior to copying them to other GPUs in a multi GPU scenario. For --split-mode row this has made a significant difference, though the quantization is still duplicated. It's still unclear though how much benefit there would be for a proper parallelization of multiple GPUs.

[am17an]

Yes, we can re-use the same structure for copying over quantizations as well. Some perf numbers (5090)

Model	Test	t/s c8a3798	t/s cuda-cache	Speedup
gpt-oss 20B MXFP4 MoE	pp2048	14084.00	14407.10	1.02
gpt-oss 20B MXFP4 MoE	pp4096	14023.45	14320.17	1.02
gpt-oss 20B MXFP4 MoE	pp8192	13704.75	14011.17	1.02
qwen3moe 30B.A3B Q4_0	pp2048	7856.41	8047.80	1.02
qwen3moe 30B.A3B Q4_0	pp4096	7663.71	7846.45	1.02
qwen3moe 30B.A3B Q4_0	pp8192	7283.03	7444.40	1.02

[am17an]

Add the cache to mmvq helps in TG as well

Model	Test	t/s c8a3798	t/s cuda-cache	Speedup
gpt-oss 20B MXFP4 MoE	tg128	389.55	396.06	1.02
mistral3 14B Q8_0	tg128	100.54	101.12	1.01
qwen3moe 30B.A3B Q4_0	tg128	325.67	332.69	1.02