Gradient implementation for Solve op

Fix gh-4585. Cast the constant to the right dtype.

keep stack trace: fix optimizations or tests (tensor/tests/test_opt.py)

test_local_adv_sub1_adv_inc_sub1.test_stacktrace: refactored and fixed unit test and removed code that was never executed, added missing copy_stack_trace call in the opt.

Pre-instantiating GPUOps

[nit] Fix typo in tensor/basic.txt: lacks a space

Updates and cosmetic changes for config doc

Symbolic Cholesky gradient implementation

Opt related changes.

Better error message

Fix crash during profile optimizer

In get_scalar_constant_value, do not return internal memory of constant

Fix a dependency version to fix travis error

Update doc with instructions for using new gpu backend

Make recent scan opt change deterministic.

Pool shape

TensorInv op added, test for tensorinv added

libgpuarray api changes