Correct dealiasing in minimal_callsafe_copy

[pablosanjose]

Closes #267

We use the approach described in that issue: redefine minimal_callsafe_copy methods throughout Quantica to restore aliasing relations between subparts of objects, particularly between GreenSolvers and any parent ph::ParametricHamiltonian.

We also fixed a silly bug that called call! on each ph[hybrid(dn)] indexing, so this might even improve performance a bit. I didn't profile any potential allocations regressions of the PR. The changes are subtle and hence a bit risky.

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 96.00000% with 3 lines in your changes are missing coverage. Please review.

Project coverage is 92.41%. Comparing base (6ab78f6) to head (49a538b).

Files	Patch %	Lines
src/types.jl	82.35%	3 Missing ⚠️

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #269      +/-   ##
==========================================
- Coverage   92.44%   92.41%   -0.03%     
==========================================
  Files          37       37              
  Lines        6061     6091      +30     
==========================================
+ Hits         5603     5629      +26     
- Misses        458      462       +4     

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

[pablosanjose]

A twist to this plot:

In a g::GreenFunction with g.solver::SparseLUGreenSolver, the solver need to alias it's parent g.contacts, because it incorporates the contact self energies into the factorization problem (instead of adding it after the fact using a T-matrix approach). Hence, if we create g, and then do g´ = minimal_callsafe_copy(g), the link between the two is broken. To restore it we need to pass not only parentham but also contacts to minimal_callsafe_copy(g.solver, parentham, contacts).

MWE

julia> glead = LP.linear() |> @onsite((; o = 1) -> o) - hopping(1) |> greenfunction(GS.Schur(boundary = 0));

julia> g = LP.linear() |> -hopping(1) |> supercell |> attach(glead) |> greenfunction;

julia> g´ = Quantica.minimal_callsafe_copy(g);

julia> g´(0, o = 0)[]  # should be a finite number, but contacts don't update anymore
1×1 OrbitalSliceMatrix{Matrix{ComplexF64}}:
 NaN + Inf*im

[pablosanjose]

So from this the following general rule is extracted that clarifies this whole PR.

The allowed aliasing relations are strictly the following:

• A AppliedGreenSolver may alias a parent AbstractHamiltonian and a parent Contacts, precisely those in the GreenFunction that contains that solver. Hence, minimal_callsafe_copy(::AppliedGreenSolver, ...) needs both of these as extra arguments in general.
• An AbstractSelfEnergySolver cannot alias anything outside of itself. Hence minimal_callsafe_copy(::AbstractSelfEnergySolver) is single-argument.

This begs the question: why shouldn't we move g.contacts and g.parent into g.solver for each AppliedGreenSolver? That would simplify minimal_callsafe_copy, which would become single-argument always. Ah, but not really, since this move would not remove the aliasing relations between deeply nested parts of the solver, such as the SchurFactorSolver and the AppliedSchurGreenSolver, so some multi-argumen minimal_callsafe_copy would have to remain.