refactor(llvm): Complete Call instruction modularization + Phase 21 organization

## LLVM Call Instruction Modularization
- Moved MirInstruction::Call lowering to separate instructions/call.rs
- Follows the principle of one MIR instruction per file
- Call implementation was already complete, just needed modularization

## Phase 21 Documentation
- Moved all Phase 21 content to private/papers/paper-f-self-parsing-db/
- Preserved AI evaluations from Gemini and Codex
- Academic paper potential confirmed by both AIs
- Self-parsing AST database approach validated

## Next Steps
- Continue monitoring ChatGPT5's LLVM improvements
- Consider creating separate nyash-llvm-compiler crate when LLVM layer is stable
- This will reduce build times by isolating LLVM dependencies

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <noreply@anthropic.com>

src/backend/llvm/compiler/codegen/instructions/mod.rs

@@ -10,6 +10,7 @@ mod strings;
 mod arrays;
 mod maps;
 mod arith_ops;
+mod call;
 
 pub(super) use blocks::{create_basic_blocks, precreate_phis};
 pub(super) use flow::{emit_branch, emit_jump, emit_return};
@@ -20,3 +21,4 @@ pub(super) use arith::lower_compare;
 pub(super) use mem::{lower_load, lower_store};
 pub(super) use consts::lower_const;
 pub(super) use arith_ops::{lower_binop, lower_unary};
+pub(super) use call::lower_call;