fix: make hot loop alloc-free and test it

[gdalle]

• Add a few tweaks to remove all allocations from the hot loop of PDHG and PDLP:
  • Move progress meter creation to initialization time
  • Remove string formatting
  • Replace homemade proj_box with built-in clamp (not sure why it helps)
• Add allocation tests using Chairmarks (note that those can fail in CI due to code coverage, hence the conditional toggle in Test.yml)

[klamike]

LGTM except I'd rather not have prog as part of the state. What do you think of https://github.com/JuliaDecisionFocusedLearning/CoolPDLP.jl/pull/51/changes#r3197969644

[klamike]

I'll try to reproduce the 1.10 CUDA failure to see what's happening. It might be too old Julia + too new CUDA (v13)

[gdalle]

Merging this since I reverted the only change you disagreed with! Thanks for the review