Adapt local_gpu_elemwise optimization of new gpuarray back-end to avoid number of inputs overflow with Elemwise<add,mul>. The current optimization was already splitting the input, but it was not using the method split_huge_add_or_mul because of new gpuarray lifter signature (see comment 
https://github.com/Theano/Theano/pull/5852#discussion_r114145523)

The unit test for large number of inputs was invalid because it was testing the old back-end (theano.sandbox.cuda). It is now adapted to gpuarray lifter optimization function. The number of settings tested is reduced to lower the computation time while still making sure we test at least one case with no number of inputs overflow and at least one case with number of inputs overflow.

split_huge_add_or_mul() is made more general so it can be used if any case like Elemwise<add,mul> occurs elsewhere.

BUG: fix bytes/string concat error, closes #5887

Remove bug in optimization

[BUG, CRASH] Fixes in DebugMode for GPU

The addition gave non-sensical results.

It also fixes a race condition.

CPU wraps around, but GPU returns max_int16, so DebugMode was
complaining.

Pass the right variable when not working inplace.

The corresponding CPU test does the same

[BUG,CRASH] Fix crash in GPU advanced indexing with no arrays

minor fix to test_nnet.py

Add sigmoid_binary_crossentropy function

Add magma gpu svd and matrix inverse

Update libgpuarray version in the doc