M5: extract render_compute_function to free functions; drop friend CodeEmitPass

[petlenz]

Addresses M5 from tracking issue #48 (multi-lens review of stack #44-#47).

What

Promotes the function-frame rendering helpers (`render_compute_function`, `param_decl`, `output_decl`, `tensor_arg_count`) from private members of `ConstitutiveModel` to free functions in a new `include/numsim_codegen/recipe_render.h`. Drops the `friend class CodeEmitPass;` declaration.

Why

The friend declaration was a one-way ratchet: every future emit pass (`TangentEmitPass`, `MoosePropertyEmitPass`, `StateVarEmitPass`, `AbaqusEmitPass`) would either need to be friended on `ConstitutiveModel` or route through `CodeEmitPass`. O(passes) friends on the user-facing recipe class.

The render helpers only needed `ConstitutiveModel`'s read-only public accessors (`name()`, `symbols()`, `outputs()`) — they never touched private state. Friend access was leakage from the original single-emitter monolith. Promoting to free functions is a pure refactor with no API change for users.

Why M5 first in the issue #48 sequence

Pure refactor, no API change, smallest blast radius. Unlocks M4 (`RecipeView` to decouple `PassContext` from raw `ConstitutiveModel`) by establishing the public-accessor-only contract that emit passes use.

Verification

• 105/105 tests pass, no regressions
• No new tests needed — pure refactor; existing tests exercise the call path through `CodeEmitPass` (`CodeEmitPass.ProducesNonEmptyComputeFunctionSource` in particular)
• Generated source is identical (same statement ordering, same template parameter naming)

Files

• `include/numsim_codegen/recipe_render.h` (new) — declares `render_compute_function` as the public-API surface for emit passes
• `include/numsim_codegen/recipe.h` — removes 4 member functions + `friend class CodeEmitPass`; adds inline definitions of the free functions in `namespace numsim::codegen` (with private helpers in `detail::`); updates `CodeEmitPass::run` to call the free function

Sequence

Per #48 follow-up plan: M5 → M4 → M3 → M2 → M6, all stacked on `main`. Each will be a small focused PR with verification.