CUDA: Hide latency of bias and gate-loading for fused mul_mat_vec_q

[ORippler]

This PR hides latency of bias and gate-loading for fused mul_mat_vec_q kernel by loading them into registers before computation of the dot-product, effectively batching them together with said dot-product. As a lot of threads are alive in this part of the kernel still, the warp scheduler has enough threads available to effectively hide the cost of loading those two single floats.

This gives 3-14% E2E speed-up for gpt-oss models (qwen3moe does not use bias and gate, and I am unaware of any other MoE model that uses bias and gate which I could run E2E perf tests on). The kernel themselves are up to 20% faster for gpt-oss.

GPU	Model	Test	t/s master	t/s this branch (babfd19)	Speedup
RTX 4000 SFF Ada	gpt-oss 20B MXFP4 MoE	tg128	85.41	88.20	1.03
RTX 4000 SFF Ada	gpt-oss 20B MXFP4 MoE	tg256	85.66	88.57	1.03
RTX 4000 SFF Ada	gpt-oss 20B MXFP4 MoE	tg512	84.80	87.62	1.03
RTX 6000 Ada	gpt-oss 20B MXFP4 MoE	tg128	240.55	248.09	1.03
RTX 6000 Ada	gpt-oss 20B MXFP4 MoE	tg256	245.03	253.23	1.03
RTX 6000 Ada	gpt-oss 20B MXFP4 MoE	tg512	242.32	250.41	1.03
RTX PRO 4500 BW	gpt-oss 20B MXFP4 MoE	tg128	212.97	223.10	1.05
RTX PRO 4500 BW	gpt-oss 20B MXFP4 MoE	tg256	220.06	236.34	1.07
RTX PRO 4500 BW	gpt-oss 20B MXFP4 MoE	tg512	218.21	237.13	1.09
RTX PRO 6000 BW Max-Q	gpt-oss 120B MXFP4 MoE	tg128	198.64	208.72	1.05
RTX PRO 6000 BW Max-Q	gpt-oss 120B MXFP4 MoE	tg256	222.68	238.11	1.07
RTX PRO 6000 BW Max-Q	gpt-oss 120B MXFP4 MoE	tg512	224.01	240.44	1.07
RTX PRO 6000 BW Max-Q	gpt-oss 20B MXFP4 MoE	tg128	296.58	338.69	1.14
RTX PRO 6000 BW Max-Q	gpt-oss 20B MXFP4 MoE	tg256	314.90	356.62	1.13
RTX PRO 6000 BW Max-Q	gpt-oss 20B MXFP4 MoE	tg512	329.31	353.40	1.07

[TinyServal]

Fixes #16815, benchmarks for affected devices (ampere cards with low memory bandwidth) can be found in the comments.

Results on the RTX A4000:

Model	Test	t/s master (b9ce940)	t/s #16847 (babfd19)	Speedup
gpt-oss 20B MXFP4 MoE	tg128	113.70	118.39	1.04
gpt-oss 20B MXFP4 MoE	tg256	112.42	116.77	1.04
gpt-oss 20B MXFP4 MoE	tg512	109.96	114.29	1.04