[AUTOTVM] Automated Operator Optimization

[merrymercy]

This is the tracking issue for bringing auto-tuning module to tvm. The auto-tuning module is critical to the performance and portability of tvm.
Since this is a big change, I will send several PRs to complete the integration. Here is the roadmap.

Step 1: Basic auto tuning infrastructure (#1312)

The core components to support auto-tuning in tvm:

• Template and space definition
• Measurement executor
• Basic tuners: random tuner, gridsearch tuner, xgboost tuner
• Tuning records/log
• Tutorial on writing tunable schedule

Step 2: Rewriting TOPI with autotvm template (#1487)

• Rewrite performance-bottleneck operators (e.g. conv2d) in tunable style
  Currently we store parameters of schedule on source file. After this step, we will release a
  log file which stores the parameters for common shapes.
• Add automated space generation for some backends
• Tutorial

Step3: Tuning a whole NNVM graph

• Tune a whole nnvm graph with easy API
• Tutorial

[tqchen]

Close this as most two steps are done