Tensorkrowch

This release fixes some things regarding the new features added in releases v1.1.0 and v1.1.1:

1. Add `log_scale` argument to `norm` of `MPS`, to return the norm in logarithmic scale.
2. Allow to apply `renormalize` to `StackNodes`, returning also `StackNodes`.
3. Add `tensors` method to `MPS` and `MPO`, to recover the list of all their tensors.
4. Add `update_bond_dim` to update the bond dim of `MPS` and `MPO` when this is changed.
5. Remove the need to set `rank` in `canonicalize` of `MPS` and `MPO` when one wants to set as rank the previous `bond_dim`.
6. Fix renormalization factors in `canonicalize` and `entropy`.

Also, we have added a new section with example notebooks, and an automated test workflow.

This release solves many bugs of release v1.1.0, and adds some new features that should have been in v1.1.0:

1. Change `unit` initialization of `MPS` by `canonical`, and add a new `unit` initialization.
2. Add operation `div`.
3. Add operation `renormalize`.
4. Add argument `renormalize` in `contract` and `partial_density` of `MPS`. Also in `contract` of `MPO`.
5. Change `mi` by `entropy` in `MPS`.
6. Add `canonicalize` to MPO.

The new release adds many new features, mainly affecting the MPS classes:

1. All MPS classes have been restructured. Now, `UMPS`, `MPSLayer` and `UMPSLayer` inherit from `MPS`. Also, the convolutional versions (`ConvMPS`, `ConvUMPS`, `ConvMPSLayer` and `ConvUMPSLayer`) inherit from the respective MPS classes and a new `AbstractConvClass` which defines the convolutional version of the `forward` method.
2. All classes can be instantiated with a given list of `tensors`.
3. The `MPS` class is now more general. One can specify which nodes are used as input, and the rest will be used as output. This is specified via `in_features` or `out_features`. This allows to pass input data with different number of features.
4. Also, the `contract` method is much more flexible. It allows to marginalize all output nodes that have not been contracted with input data. Also, in between this MPS-MPS contraction that is performed when marginalizing, one can add `embedding_matrices`, or an `mpo`.
5. Canonical forms in MPS allow to `renormalize`.
6. Methods `copy` and `parametrize` have been overriden to work better for MPS classes.
7. New methods have been added: `norm`, `partial_density` and `mi`.
8. New class `MPSData` that do not inherit from `MPS` has been added.
9. New classes `MPO` and `UMPO` have been added. These can be used to be contracted with input data nodes, with a `MPSData`, or used as input to a `MPS`.
10. `initialize` method has been extended for all `MPS` classes and `MPO`. For `MPS`, the methods `"randn_eye"` and `"unit"` have been included, in addition to the usual initialization methods of nodes.
11. New embedding functions have been added: `discretize` and `basis`.
12. New submodule `decompositions` has been added, with functions to turn vectors into MPSs (`vec_to_mps`), and matrices into MPOs (`mat_to_mpo`).
13. `StackNodes` can now be reconnected via `stack_1 ^ stack_2`, instead of having to reconnect all their edges.
14. All functions derived from or that call `split`, allow for passing some or all of the arguments `rank`, `cum_percentage` and `cutoff`, instead of only one of them.
15. In-place operations for edges have got their respective non in-place operations, including `contract`, `svd`, `svdr`, `qr` and `rq`.
16. New methods for nodes: `numel`, `change_type` and `is_connected_to`.
17. New methods for `TensorNetwork`: `delete`.
18. `reattach_edges` method of nodes now admits a list of axes, to reattach only the selected edges.
19. The reserved name `"virtual_stack"` has been replaced by `"virtual_result"`.

The C++ extension has been eliminated in favor of pure Python optimizations. These optimizations primarily impact the following operations:

- contract_edges
- contract_between
- stack
- unbind
- einsum

First release of TensorKrowch!