relax the error constraints on these tests due to likely
FP compuation accuracy issues.

Add trace call expr to allow trace Tensor data
at the runtime. By default the default handler
is enabled which prints a tracing data to stdout,
otherwise user should specify a call_back as
global_function (aka @tvm.register_func).

The issue is related to:
https://discuss.tvm.ai/t/idea-trace-expression/945

* [Relay][Keras] force const dtype to be float32

* fix pylint

* Enhanced simplification rules for Div by a positive constant

* Fixed my last commit to correctly interpret TVM's division as truncated division

* Fixed implemenation of IntSet::can_prove_non_positive()

* addressed comments by @yzhliu

* addressed comments by @sgrechanik-h

* addressed more comments by @yzhliu

* Fix clock type in rpc_session timer

* Fix lint error

* [TOPI] Add roi align

* Refactor bilinear in image resize

* Rename to roi_align_nchw

* Fix

* Add cast op
* Rename dtype_cast to cast
* Add additional safety check for String2TVMType
* Add missing relay op docs

dtype of count is the same as dtype of inputs[0] when created, but its type may
  change when multiplied by inputs[0]->shape[i]. Which causes dtype of
  output is not same as dtype of input.