warn on `batched_mul_generic!`?

[chengchingwen]

The batched_mul_generic! is really slow on both cpu and gpu, but somehow it isn't something easy to find. It might worth displaying a one time warning like the cuda scalar indexing warning.

[ToucheSir]

Between this and #475 I wonder if we shouldn't have a more general "slow fallback" policy like other libraries do. The simplest version of this would be toggleable warnings when generic methods are called. Could this be generalized and/or made more sophisticated?

[mcabbott]

The slow paths for conv do print warnings, so batched_mul! could at least follow that.

Agree that ideally there would be a way to make these errors, like CUDA. (Whose default is now error not warning.)

Xref FluxML/Flux.jl#2454, about adding some similar switch.