slightly relax batch norm check

[bwilbertz]

As already raised in #4704, the current LayerSetup method of the BatchNormLayer expects to have no LayerParamter configured at all, but also adds 3 LayerParameter itself (cf. batch_norm_layer.cpp).
This all is done to avoid any unintenional backward modification of the global stats in the BatchNormLayer and therefore forcing lr_mult = 0.f.

As a consequence, a valid (upgraded) net, which was passed once into Net<Dtype>(net_param) and read back via ToProto(...) cannot be passed a second time into Net without being upgraded again.

This PR slightly relaxes the checks in LayerSetUp, to just ensure lr_mult == 0.f or adding it, if the LayerParameter is missing (as it was introduced in #4704).

[shelhamer]

Thanks for issuing a patch for this @bwilbertz and sorry I couldn't handle it immediately. I'll try to check and merge today.

Note that the CI failures seem to be some bogus issue with CUDA packages on the Travis machines: https://travis-ci.org/BVLC/caffe/jobs/163831672 We can hopefully resolve that shortly. @lukeyeager have you seen this issue w/ Travis?

[lukeyeager]

@lukeyeager have you seen this issue w/ Travis?

Yeah we're working on it - sorry.

[shelhamer]

Thanks again for this @bwilbertz -- merged after checking it myself since the CI issue is unrelated.

@lukeyeager thanks for looking into the CI problem!