Thread no_recursion through compile_program for the FFI

`Compiler::compile_program()` was the only `LLVMJIT::new()` site that
forgot to copy `self.no_recursion` onto the JIT before calling
`compile_only`, so the call-depth prologue/epilogue was still emitted
on every function regardless of the compiler-level flag.

That codepath is what `lyte_compiler_compile` calls, which is the only
LLVM entry the xcframework uses, so flipping the FFI default to
`no_recursion = true` in 2df4d9d had no effect on the actual LLVM
backend — Audulus DSP code was still paying the full ~2x reverb
regression documented in docs/LLVM_CALL_DEPTH_REGRESSION.md. The CLI
(`run_llvm` / `print_llvm_ir`) was unaffected because those sites do
thread the flag, which is why the bug was invisible to CLI smoke tests.

Verified end-to-end with a C harness against `liblyte.dylib` driving
`lyte_compiler_compile` on the regressed reverb script: 10000 process()
calls drop from ~0.43s to ~0.21s once the line is added.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: wtholliday

src/compiler.rs

@@ -878,6 +878,7 @@ impl Compiler {
         {
             let mut jit = crate::llvm_jit::LLVMJIT::new();
             jit.print_ir = self.print_ir;
+            jit.no_recursion = self.no_recursion;
             let llvm_prog = jit.compile_only(&self.decls, &entry_points)?;
             return Ok(CompiledProgram::Llvm(llvm_prog));
         }