Experimental Conv2D implementation specialized for deep convolutions (i.e....

Experimental Conv2D implementation specialized for deep convolutions (i.e. large in_depth * out_depth, see cost model in deep_conv2d.cc for details).
Can be enabled/disabled with environment variable (disabled by default).
Currently only supports 3x3 filter transforms, but more transforms (and performance work) to come.

// CPU 1-thread
Benchmark                          Base (ns)  New (ns) Improvement
------------------------------------------------------------------
BM_ConvFloatFwdCPU1_conv04          130679850  90340598    +30.9%
BM_ConvFloatFwdCPU1_conv13          133216209  97856032    +26.5%
BM_ConvFloatFwdCPU1_conv23          153311103  112391095    +26.7%
BM_ConvFloatFwdCPU1_conv28          111367333  84726090    +23.9%
BM_ConvFloatFwdCPU1_conv38          61715027  50013699    +19.0%
BM_ConvFloatFwdCPU1_conv43          100892227  92104830     +8.7%
BM_ConvFloatFwdCPU1_conv48          156439547  130304041    +16.7%
BM_ConvFloatFwdCPU1_conv52          769133078  647485239    +15.8%

// CPU 16-threads
Benchmark                          Base (ns)  New (ns) Improvement
------------------------------------------------------------------
BM_ConvFloatFwdCPU16_conv04         219668015  175358604    +20.2%
BM_ConvFloatFwdCPU16_conv13         199494286  129051236    +35.3%
BM_ConvFloatFwdCPU16_conv23         227566760  145085849    +36.2%
BM_ConvFloatFwdCPU16_conv28         162253350  107485116    +33.8%
BM_ConvFloatFwdCPU16_conv38         85966024  61742460    +28.2%
BM_ConvFloatFwdCPU16_conv43         152369560  108266340    +28.9%
BM_ConvFloatFwdCPU16_conv48         235960229  158102948    +33.0%
BM_ConvFloatFwdCPU16_conv52         1082339398  810291519    +25.1%
Change: 131850132