refactor: immutable BuiltModel bundle replacing the mutable DOCP cache

[ocots]

Summary

Replaces the mutable backend cache in the DOCP flow with an immutable BuiltModel.

build_model previously returned a bare NLP and (for the Exa backend) mutated
exa_getter into the shared DiscretizedModel.cache so that build_solution
could reuse it. The getter is a co-product of building the model and cannot be
cheaply recomputed. This PR makes build_model return an immutable bundle that
carries both the NLP and any build-time auxiliary, so nothing is mutated.

Changes

• Optimization: new BuiltModel{problem, nlp, cache} + NoCache. build_model
  returns a BuiltModel; build_solution(built, sol, modeler) dispatches on it.
  The orchestrator threads built.nlp into the solver; the mid-level
  solve(nlp, solver) stub stays untyped (NLPModels weak dep).
• DOCP: nlp_model returns the bare NLP (build_model(...).nlp);
  ocp_solution realigned onto BuiltModel.
• Solvers: extract_solver_infos moved here from Optimization (solver-side utility).
• Deps: KernelAbstractions promoted to a hard dep so the CTSolversExaModels
  extension triggers on ExaModels alone (downstream packages such as CTDirect
  needn't using KernelAbstractions); Aqua stale_deps ignores it since the main
  module does not reference it directly.
• Tests: each modeler-using test file now imports ADNLPModels/ExaModels so the
  extensions trigger under per-file isolation; fixtures wrapped in BuiltModel.

Why it matters

• Fully immutable cache; the Exa getter is never recomputed (zero overhead).
• Removes a latent bug: two build_model calls on the same DiscretizedModel
  could clobber each other's getter; a stale getter could survive an ADNLP build.

Companion

CTDirect adopts the same bundle (DOCPCache becomes immutable, getter rides an
ExaBuildCache) on its refactor/docp-dispatch branch (PR #601).

Tests

Full CTSolvers suite green (non-extension suites + extensions, ~3100 tests),
including Aqua.

🤖 Generated with Claude Code