Fix extra allocations in get_gradient!'s jacobian

[Affie]

I'm not sure of the correct way to fix this, but here is the cause of the extra allocations in my example from: JuliaRobotics/IncrementalInference.jl#1546 (comment)

With this fix it shaves off 15 seconds of a 30-second solve.

[Affie]

This is where all the allocations come from, in my case, I give a sparse jacobian as input.

[mateuszbaran]

That's a neat solution 👍

[codecov]

Codecov Report

Merging #300 (c552eb3) into master (48f5ab0) will decrease coverage by 0.25%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #300      +/-   ##
==========================================
- Coverage   99.88%   99.64%   -0.25%     
==========================================
  Files          77       77              
  Lines        7083     7083              
==========================================
- Hits         7075     7058      -17     
- Misses          8       25      +17     

Files	Coverage Δ
src/plans/nonlinear_least_squares_plan.jl	85.71% <ø> (-14.29%)	⬇️
src/solvers/LevenbergMarquardt.jl	100.00% <100.00%> (ø)

... and 2 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[kellertuer]

To summarise a lengthy chat.

The idea proposed here breaks with one central design idea, that the problem (manifold, cost function , gradient,...) is stored in one struct and the state (everything the solver needs between iterates) in another.
It also abstracts accessors to the problem from implementation details.

With these ideas the current (though short) PR breaks – quite a bit.
I do not yet see a good way to resolve this, sadly, but I acknowledge, that “passing state memory to the problem/objective” might be something useful to have.
Since this summarizes/conclucdes my code remarks above, I marked them resolved (though they are technically not).

[mateuszbaran]

I've submitted an alternative fix that doesn't break Manopt's design: #301 -- it should also fix this issue.

[kellertuer]

Superseeded by #301