Improvements to `alternating_update`

[mtfishman]

Here's an issue to track improvements to make to the general tensor network solver code, alternating_update, following up on #59.

• Switch default tdvp solver to KrlyovKit.exponentiate.
• Define default_projected_operator which generically defines how to convert an input operator to a projected operator, for example default_projected_operator(x::TTN) = ProjTTN(x), default_projected_operator(x::TTNSum) = ProjTTNSum(x), etc. This could also help with creating certain caches/projected operators when solving different kinds of problems.
• Rethink the interface for choosing between different solvers as well as custom ones in tdvp, ideally we don't hard code a list of them (i.e. deprecate the solver_backend="exponentiate"/solver_backend="applyexp" interface) and instead make it easy for users to pass a solver function and solver keyword arguments.
• Remove the time argument t from alternating_update, since it only makes sense for solvers that implement time evolution like tdvp.
• Decide on the interface for tdvp. How to specify total time, time step, number of steps, which argument ordering?
• nsweeps vs. nsteps. nsteps makes a bit more sense for a time stepping algorithm like tdvp, but nsweeps is more generic (i.e. defines a sweep through the graph). Also a step can involve multiple sweeps in a higher order method.
• Improve keyword argument code patterns based on: https://itensor.github.io/ITensors.jl/dev/DeveloperGuide.html#Keyword-Argument-Best-Practices
• Generalize ProjTTN types to a more general ITensorNetworkCache type (or specifying custom caching types which are relevant for different contraction backends), based on a contraction sequence tree. Also allow custom contraction sequences and contraction backends, to generalize to optimizing/updating other tensor networks.
• Replace ProjTTNSum with a more general lazy sum, based on the Applied type in ITensors.LazyApply.
• Generalize to accept AbstractITensorNetwork, and ensure the interface requirements are general enough to work for more general networks.
• Think more carefully about the interface and implementation for solving various linear equations, like Ax ≈ λx, Ax ≈ λBx, Ax ≈ b, x ≈ y, etc. which shares caching structures, works with general tensor networks, contraction backends, etc.

[emstoudenmire]

• Switch default tdvp solver to exponentiate

(Feel free to move this up – I just can't edit your list above.)

[mtfishman]

Thanks, I added it.

[emstoudenmire]

Proposed redesign of internals:

1. Eliminate outermost for loop over substep (see the function update_step in update_step.jl)
2. Pass a data structure which says how to do a single (full) sweep over the system. This defines what a sweep means, as in nsweeps=3 does three of these.
3. This data structure can carry an optional "payload" (named tuple?). Using tdvp as an example, the payload would store the explicit, signed and weighted time steps, already preprocessed. In the case of a PEPS it could store the row and column number.
4. There is an optional, second type of callback function (not a solver) which can post process the payload before feeding it into the solver. (May not be necessary to add this callback feature if the payloads are sufficiently preprocessed.)

Basically Lander's data structures become the definition of what a sweep is and are generalized to not just store the forward/reverse sign, but any data relevant to that specific point in the sweep.

A related part of this redesign is that the observer system will handle all printing, via custom "printing callbacks" which are appending to the user-provided list of observer functions (this is in addition to 1-4 above).