Add solve and svn to dispatched function

[Ericgig]

Description
Add data layer functions needed to update steadystate to use qutip's data layer.

Most of steadystate's methods solve the system L(rho_ss) = 0 using different solver from numpy, scipy, mkl.
This add the solve dispatched function to wraps this operation at the data layer level.
For the CSR version, it gives access to all methods available in scipy.sparse.linalg and our mkl version.
For the Dense version, solve and lstsq from numpy.linalg are made available.
Options supported by numpy / scipy for each function are pass through:
qutip.data.solve(A, B, method, options) call scipy.sparse.linalg.method(A, B, **options).

Other methods of steadystate decompose the Liouvillian with either eigen or svd.
svd was added as a dispatched function, but only with specialization for Dense.
I made a sparse version, but svds can't readily compute all singular values and does not respond well to splitting the problem into 2 like it was done for sparse eigensolver.

This PR only include the new dispatched functions, not updated steadystate using them or Qobj interface.

[hodgestar]

I think allowing specializing solving and SVD is extremely useful. I have a few questions though:

• Do you have a specific use case that you're adding it for right now?
• It will be hard to support methods such as "mkl_spsolve" on a strange data backend, so essentially inside qutip itself it is not okay to pass a method except when it was supplied to us by the user. Do you have any idea how to ensure we don't forget that rule? Or any ideas on how to provide an interface that allows specifying a method but which also doesn't break if someone hardcodes a method name in a place that is difficult for users to override?

[Ericgig]

svd is used by steadystate.

There are no method in common between Dense and CSR, the default is solve for Dense, but spsolve for CSR. All iterative methods are only available in sparse.linalg... So we cannot pass it anywhere thus I think it won't be hard to remember.

If we want to have the default change depending on whether we have mkl or not, we could read the default method from core's settings.