Fix BatchNorm crash under Mooncake in eval mode

[AstitvaAggarwal]

While running BatchNorm through Mooncake in eval mode, the backward pass crashes with:

MethodError: no method matching unsafe_convert(::Type{CuPtr{Nothing}}, ::Nothing)

The cause: cudnnBNForward! only fills in the cache's mean/ivar when training=true — in eval mode it just leaves them unset. But cudnnBNBackward! doesn't check for that; it assumes any non-nothing cache is already populated and hands whatever's in it straight to cuDNN.

The Mooncake rule for batchnorm always builds a live cache, whether it's training or not, so eval-mode calls hit this every time. This is easy to trigger via Flux.BatchNorm, since its training/eval auto-detection (NNlib.within_gradient) never returns true under Mooncake — that trick only works for Zygote.

This PR makes two small fixes:

• Only build the cache when training is actually happening (matches what the "no cache" path already does safely).
• If the caller passes their own cache object, use it instead of quietly replacing it with our own.

I tested this directly on GPU: reproduced the crash on the released NNlib, confirmed it's gone with the patch, checked the gradient still matches Zygote, and ran through all four combinations of training/track_stats to make sure nothing else was missed.

[AstitvaAggarwal]

@CarloLucibello I think this is good from my side, please let me know if this needs anything more. Thank you