Consider merging `cuda` and `cudnn` features

Currently, `cuda` and `cudnn` are separate features, that implement many of the same kernels in different ways.
This results in duplicated effort on candle's part, more configuration flags, more kernels, more complicated control flow in candle - all to support `cuda` without `cudnn` separately from `cuda` with `cudnn`.

Most mainstream ML frameworks (pytorch, tensorflow, jax) just assume cuda + cudnn as a dependency together, avoiding the duplicated effort.

I suggest candle should merge `cudnn` and `cuda` configuration flags together, and always require `cudnn` on `cuda` hosts.