Add support for Jacobians

[adrhill]

This PR adds fallback implementatios for value_and_jacobian!, value_and_jacobian, jacobian!, and jacobian.

The implementation dispatches on AbstractForwardBackend and AbstractReverseBackend:

• AbstractForwardBackend: Makes use of value_and_pushforward!, which computes JVPs. The pushforward of a standard basis vector $e_i$ computes the $i$-th column of a Jacobian. The Jacobian is therefore assembled column-wise.
• AbstractReverseBackend: Makes use of value_and_pullback!, which compute VJPs. The pullback of a standard basis vector $e_i$ computes the $i$-th row of a Jacobian. The Jacobian is therefore assembled row-wise.

This requires the implementation of a standard basis vector in src/unitvector.
This implementation can be replaced with more performant implementations in the future by specializing on backend types.

To write into pre-allocated Jacobians, type annotations of pullbacks and pushforwards had to be loosened.

Towards #6.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 85.07463% with 10 lines in your changes are missing coverage. Please review.

Project coverage is 88.82%. Comparing base (c65fb27) to head (538fb6b).

Files	Patch %	Lines
src/jacobian.jl	87.23%	6 Missing ⚠️
src/forward.jl	50.00%	2 Missing ⚠️
src/reverse.jl	50.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #22      +/-   ##
==========================================
+ Coverage   85.98%   88.82%   +2.84%     
==========================================
  Files           9       11       +2     
  Lines         107      170      +63     
==========================================
+ Hits           92      151      +59     
- Misses         15       19       +4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[adrhill]

Good news: only Enzyme required specialized unitvector and value_and_pullback! methods to support inplace mutations of Jacobians!

In both cases, this was due to Enzyme's Duplicated(x::T, dx::T) assuming identical types for inputs and outputs of the pullback.

---

One annoyance encountered in this PR is that value_and_pullback!(dx, ..., x::Real) assumes that dx is scalar and can't be mutated, even though it might in fact be a mutable entry in a pre-allocated Jacobian. A similar case exists for pushforwards on functions with scalar outputs y. This required an additional two methods for _value_and_jacobian!:
https://github.com/gdalle/DifferentiationInterface.jl/blob/1e40a0d01a76eb3c78141fffbad59bf9bf83add1/src/jacobian.jl#L41-L59

[adrhill]

Superseded by #23.