[Metal] Fix depthwise conv 1D kernel name for large variant

[Brooooooklyn]

The kernel name was constructed with the wrong argument order, producing depthwise_conv_1d__largefloat16 instead of depthwise_conv_1d_float16_large which matches the Metal kernel registration.

[zcbenz]

Thanks for fixing this! What was happening when calling the kernel, did it always get JIT compiled?

[Brooooooklyn]

What I got:

❯ node index.js

[MLX] Exception in async_eval: [metal::Device] Unable to load kernel depthwise_conv_1d__largebfloat16

[MLX] Exception in mlx_qwen35_moe_eval_token_and_caches: [metal::Device] Unable to load kernel depthwise_conv_1d__largebfloat16

[MLX] Exception in array_eval: [metal::Device] Unable to load kernel depthwise_conv_1d__largebfloat16

[1] 32912 segmentation fault  node index.js