Session B: codegen covers locals, branches, loops, recursion

Expand omnimcode-codegen's lower_function to handle the full pure-i64
subset of OMC bytecode. After this session, any OMC fn whose body
uses only i64 values, locals, conditionals, while loops, and
recursive self-calls JITs to native code with byte-identical
results to the tree-walk interpreter.

What's new:
- FunctionLowerer struct: extracts the per-fn state (block table,
  var slots, builder, stack) out of free functions
- Two-pass lowering: first identify block leaders (jump targets,
  post-terminator ops), then walk ops emitting per-block IR
- Locals via entry-block allocas: StoreVar / AssignVar / LoadVar
  use load/store on a HashMap-tracked PointerValue per name
- Control flow: Jump, JumpIfFalse, JumpIfTrue lower to LLVM
  conditional/unconditional branches
- Comparisons: Eq, Ne, Lt, Le, Gt, Ge produce i64 0/1 (z-extended
  from i1) so they match the bytecode VM's semantics
- Arithmetic: Div, Mod (signed), Neg added to Session A's Add/Sub/Mul
- Bitwise: BitAnd, BitOr, BitXor, BitNot, Shl, Shr
- Logical: And, Or (non-short-circuit, matches compiler emission), Not
- Recursive Call: Op::Call with name == current fn name lowers to
  a self-call. Cross-fn calls still rejected — that's Session D.

JumpIfFalse-cleanup-Pop idiom: the OMC bytecode compiler emits
JumpIfFalse (which PEEKS the condition) followed by a Pop in both
the fall-through path and at the jump target. Codegen models the
branch as consume-and-jump, then marks both cleanup Pops as no-ops
via a per-fn HashSet so we don't double-pop or stack-underflow.
Same handling for JumpIfTrue.

Tests added (8 total, all passing):
- jit_max_two_args: tests if/else, JumpIfFalse
- jit_abs_single_arg: tests Neg + comparison
- jit_sum_to_n_while_loop: tests while loop, locals (s, k via
  allocas), backward Jump, comparison-driven exit
- jit_factorial_recursion: factorial(20) = 2,432,902,008,176,640,000
  exercises recursive self-Call, conditional return base case
- jit_rejects_cross_fn_call: negative case for Session D scope guard

Still NOT yet covered (deferred to subsequent sessions):
- Floats, strings, arrays, dicts, builtins (Session D)
- Cross-fn calls (Session D)
- HBit dual-band representation (Session C — packed i64x2)
- AVX-512 intrinsics, harmony monitoring opcode, predictive correction
- Closures, match, try/catch

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: RandomCoder-lab