Reduce allocations in `evaluate_policy!`

[oyamad]

Current code path

evaluate_policy! currently:

• copies cdp.interp.Phi into a dense matrix A
• loops over states and shocks
• subtracts ckron([vec(evalbase(...)) ...]) * discount * w
• factorizes A with lu
• solves with ldiv!

Possible directions

• reduce temporary allocations from evalbase, collect, and ckron
• use more workspaces during matrix assembly
• investigate whether the matrix assembly can be made more structured/in-place

(AI assisted (ChatGPT Pro 5.4) in drafting this issue)