Ensure dynamic tolerance scaling works for every gradient algorithm

[leburgel]

Missed in #413 after #410 was merged, but the fact that the custom MPSKit._updatetol implementation was specialized to FixedPointGradient meant that the new default configuration for an optimization using fixedpoint was broken when using an ImplicitGradient.

I added such a run to the example tests, to test ImplicitGradient via the public interface, and not just through raw gradient accuracy tests.

UPDATE: This exposed some issues with the corner type of the environment returned by leading_boundary when using implicit differentiation. In this case, we return the corner as a complex non-diagonal tensor map before evaluating the energy density, since the implicit differentiation approach relies on allowing complex non-diagonal adjoints of the corner. When setting reuse_env = true, the update! runs into a type mismatch between the initial environment (which will have a real diagonal corner when using a C4vEighProjector) and the corner to be updated (complex non-diagonal when used to evaluate the energy). Instead of converting the initial environment to have a complex non-diagonal corner, I opted to convert the corner type of the updated environment to a real diagonal if necessary.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
...lgorithms/optimization/implicit_differentiation.jl	90.08% <100.00%> (ø)
src/environments/ctmrg_environments.jl	72.97% <100.00%> (+1.06%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[pbrehmer]

Looks good to me!