Add matrix exponential sensitivities

[ararslan]

Scalar exponential is covered by the DiffRules integration.

[codecov]

Codecov Report

Merging #161 into master will increase coverage by 0.03%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #161      +/-   ##
==========================================
+ Coverage   97.06%   97.09%   +0.03%     
==========================================
  Files          20       20              
  Lines         749      758       +9     
==========================================
+ Hits          727      736       +9     
  Misses         22       22

Impacted Files	Coverage Δ
src/sensitivities/linalg/generic.jl	100% <100%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update a1b119b...d0f4669. Read the comment docs.

[eperim]

Only thing I'd add is either a reference to the paper in this line, https://github.com/Theano/Theano/blob/3b8a5b342b30c7ffd2f89f0e9efef601b7492411/theano/tensor/slinalg.py#L538, or to the line itself.

[ararslan]

This implementation seems only to work for real, symmetric matrices, since the eigenvalues are real. A trivial refactoring to take the real parts of the eigenvalues and vectors earlier suggests that the implementation (or at least that refactoring) incorrect when compared to finite differencing.

[ararslan]

I've made it an error for the input not to be symmetric and added a test for that case.