Add dispatcher capacity to support any type specialisation

[Ericgig]

Description
Allow the dispatcher to accept specialisations on generic Data.
This as 2 uses:

• Simplify the creation of complex function using simpler ones: inner defined in term of matmul.
• Allow us to create niche dispatched functions for some data layer without the need to be supported by all: blas' zgemm (matmul) can transpose and adjoint of each input matrix in the same operation.

The Data dispatched functions have less priority than direct specialization. They are used directly (no _constructed_specialisation) when the Data is in the inputs, but when the output data type is set, there will always be a conversion from Data to the desired type.

In this PR:

• Add dispatch on Data.
• Add direct tests for the dispatcher.
• Create kron_transpose which is kron(left.T, tright). It's commonly used for superoperator and it can be optimized for Dense and Jax by doing it in one operation.
• Add Data specialization for inner, inner_op, expect, expect_super and trace_oper_ket.
• Make zeros_like and identity_like as dispatched function so the order is kept during operation for dense.

[coveralls]

Coverage: 77.746% (+0.01%) from 77.733% when pulling 357673b on Ericgig:feature.data.dispatch into fc9c580 on qutip:master.

[coveralls]

Coverage: ?%. Remained the same when pulling 0bcd59f on Ericgig:feature.data.dispatch into fc9c580 on qutip:master.

[coveralls]

coverage: 78.118% (+0.01%) from 78.106% when pulling 932ad81 on Ericgig:feature.data.dispatch into f66b8f2 on qutip:master.

[hodgestar]

I left a few small comments, but overall this looks great.

[hodgestar]

I'm not sure what is happening with this continue statement. Perhaps we can just do:

if not output:
    ...
else:
    ... rest of the current loop without all the `if output:` checks

[hodgestar]

If __name__ has to be here, can we put something in it like __name__ = "dummy name" rather than leaving it blank which just looks a bit odd.