Moved all apt-get install commands into same RUN line as apt-get update.  There is a subtle timebomb if apt-get update is performed in a separate RUN line, due to how docker layers work. (#179)

* isolate graph building / execution scope in some unit tests (this triggered the user-after-free which was also observed from TF)

* Change descriptor::layout::TensorViewLayout to store a shared pointer to the TensorViewType rather than a reference to the TensorView

* port functions from funky passman branch

* update pass tests

* wip

* fixed

* remove obsolete test

* remove debug prints

Include cmath for 'expf' etc.

* Type propagation for convert

* Implement "convert" through VM

* Temporary commit to unconfuse git merge

* Undo temporary commit to unconfuse git merge

* Add support for some polymorphic operators (arithmetic, etc.)

* Extend to all other ops except Tuple, which is tricky (not quite sure where to get the needed type info)

* Some tidying(???) up

* Extend to handle Tuple op

* Slightly more descriptive macro name

Detect macOS and adapt sed usage in clang-format version check script; also fix small typo in err msg (#169)

Emacs likes to add hidden symlinks with names like '.#myfile.cpp' that don't actually point to real files as a locking mechanism, which cause the format check/apply scripts to fail. This tweaks the format scripts to ignore files that are symlinks (adds a -type f predicate to the find commands), which gets around this issue.

apply new .clang-format

* Implement broadcasting of scalars through VM

* Implement vector-to-matrix broadcasting through VM

* Add missing files

* Address review comments

Remove old {get,set}_{array_matrix}{_2d,} methods; update dot product to use new Eigen wrapper (#158)

Use weak_ptr for node in inputs/outputs, turn off alignment style.

* Have abs use layout.

* Use layout to get vector info

* Add format selector for mapped arrays
Convert some ops
Drop runtime.cpp since we don't expose the functions any more.

* Add vector concat

* concat matrix

* Switch more ops to simpler eigen wrapping

* Add reduce op class and type propagation

* Implement FunctionCall through VM

* Changes Function to require explicit return type at construction time

CMake: Update MKL-DNN tag and remove upstreamed patch

* cleanup pass registration code.

* Replace explicit code to do passes with pass_manager passes

* make external function work with const types internally, it should not change the graph

* Basic layouts, tuples

* Review comments.

* Add incude for size_t.

* Bop use-after-free in GetTupleElement

* Inline call to get_element_types()

* Implement vector and matrix concatenation through VM

* Address PR comments

* Comments on Concat semantics

* Remove some redundant template parameters in concat_matrix and concat_vector

* Formatting fixes in external_function.cpp

* Vector dot product implementation

* Implement matrix mult instruction in VM

* Add implementations for scalar-tensor and matrix/vector products.

* This is kind of a cheap hack for the moment, at least in the tensor
  constant case: each constant instruction carries around its value,
  and copies that to the result buffer everytime we need it. Ultimately
  we will probably want a pass that allocates space in the call frame
  for constants, similar to what we do for parameters.

* Add elementwise ops to runtime

* equals operator implementation at graph and runtime level

* "Select" operator (elementwise ?:) at graph and runtime level

* A few file renames (`notequal` -> `not_equal`, `lessthan` -> `less_than`)

* Macro wrapping boilerplate for op registration in `external_function.cpp`
  (this covers everything except `op::Parameter`).

* Forces CI to use clang-3.9