Add batch normalization op

Add stack trace for nnet

Faster merge

In debugprint, don't print the profile headers more then once and only add it if the profile isn't empty.

The merge optimization when applied with the destroy handler can be very slow.
I have one case that got speed up from 190s to 12s by this PR:

Original:
SeqOptimizer  OPT_FAST_RUN  time 190.461s for 2161/548 nodes before/after optimization
   1.280s for fgraph.validate()
   132.634s for callback
   time      - (name, class, index) - validate time
   170.167204s - ('merge3', 'MergeOptimizer', 38) - 0.000s
     MergeOptimizer
       nb fail= 7728 merged=    0 constant=    0
       time replace=170.17 validate=0.00 callback=128.92
       callbacks_time
(<theano.gof.toolbox.PreserveNames object at 0x7f46b8864310>, 0.04287838935852051)
(<theano.gof.toolbox.ReplaceValidate object at 0x7f46b33f43d0>, 0.06792092323303223)
(<theano.tensor.opt.ShapeFeature object at 0x7f46b8864890>, 0.44670534133911133)
(<theano.gof.destroyhandler.DestroyHandler object at 0x7f46b4a44810>, 1.9552459716796875)
(<theano.gof.opt.MergeFeature object at 0x7f46b8864650>, 126.08237051963806)
   10.041219s - ('canonicalize', 'EquilibriumOptimizer', 4) - 0.051s

New one:

 SeqOptimizer  OPT_FAST_RUN  time 12.338s for 2161/548 nodes before/after optimization
   1.219s for fgraph.validate()
   3.536s for callback
   time      - (name, class, index) - validate time
   4.779337s - ('gpu_opt', 'SeqOptimizer', 12) - 0.012s
     SeqOptimizer      gpu_opt  time 4.779s for 633/504 nodes before/after optimization
       0.012s for fgraph.validate()
       0.817s for callback
       4.748835s - ('gpu_local_optimizations', 'EquilibriumOptimizer', 1) - 0.012s
...
   0.002217s - ('merge3', 'MergeOptimizer', 38) - 0.000s
     MergeOptimizer
       nb fail=    0 merged=    0 constant=    0
       time replace=0.00 validate=0.00 callback=0.00

Try to fix rtd

Add requirement specific to rtd

Fix Sphinx markup in docs

Add Docker images to the installation docs

automatic gpu selection no longer works in CUDA-TK 7.0

Enables float16 gemm on gpuarray when the cuda version supports it

This reverts commit 1814576c.

Mixed2

- pycuda initialization was also affected by the bug from CUDA 7.0

- fixing an issue introduced with CUDA-TK 7.0, when automatic selection
  of free gpu by os / library no longer works. Library always selects '0',
  which leads to crash in case of 'exclusive mode'.
- this fix is inspired by Dan Povey's fix for Kaldi:
  https://github.com/kaldi-asr/kaldi/commit/6548565445167e00125848f91d7da5f3f949b2a2
- it does a loop over gpus until a free gpu is taken.

Optimize local_fill_sink

fix shape mismatch in GpuDnnPoolGrad