CUDA gather mv

[angeloskath]

Adds a CUDA implementation for gather_mv, the gather_vm is missing but using this and @zcbenz grouped mm for the sorted version we can run unquantized MoEs on CUDA.

$ mlx_lm.benchmark --model Qwen/Qwen3-30B-A3B -p 2048 -g 128
Running warmup..
Timing with prompt_tokens=2048, generation_tokens=128, batch_size=1.
Trial 1:  prompt_tps=2101.020, generation_tps=32.027, peak_memory=61.821
Trial 2:  prompt_tps=2088.784, generation_tps=32.078, peak_memory=61.821
Trial 3:  prompt_tps=2101.889, generation_tps=32.072, peak_memory=61.822
Trial 4:  prompt_tps=2095.918, generation_tps=32.046, peak_memory=61.822
Trial 5:  prompt_tps=2112.402, generation_tps=30.730, peak_memory=61.822
Averages: prompt_tps=2100.002, generation_tps=31.791, peak_memory=61.822

[awni]

Awesome!!