perf(vm): dispatcher coverage for closures, varargs, multi-return, loops, self, concat

[davydog187]

Closes #272

Summary

Extends Lua.VM.Dispatcher and Lua.Compiler.Bytecode to cover the remaining opcodes needed for closures-, OOP-, and string-heavy workloads. After this PR, the encoder's catch-all :fallback clause only matches genuinely out-of-scope opcodes (:goto/:label, :set_global, the bitwise family). The closures, oop, and string_ops benchmarks compile end-to-end with no fallback in any sub-prototype.

What's covered

New opcodes (tags 37–51, contiguous after B5b-v2):

• Closures + upvalues — :closure, :set_upvalue, :get_open_upvalue, :set_open_upvalue
• Vararg + multi-return — :vararg, :return_proto_varargs, :return_collect, :return_multi, :call_multi (handles every shape of :call outside the existing :call_one/:call_zero fast paths)
• Loops + break — :while_loop, :repeat_loop, :generic_for, :break. The B5b-v2 contains_break? guard against break inside :numeric_for is removed.
• OOP — :self, :concatenate

:tail_call was listed in the plan frontmatter but is verified dead code (codegen never emits it; only Lua.Compiler.Instruction.tail_call/3 exists as an unused helper). Tail-position calls compile to :call with result_count = -1, covered by :call_multi.

Frame tuple changes

The frame's dest slot gains a {:multi, base, count} variant for multi-return shapes (-1 forward, -2 expand, n > 1 fixed multi). :call_one/:call_zero keep their integer / :discard fast paths — the fib hot path stays on the integer-base branch. return_one/3 and return_multi/3 both pattern-match on all four shapes.

Loop CPS

:while_loop/:repeat_loop/:generic_for/:numeric_for each push two cont markers: a CPS marker (driving normal body/condition handoff) and a :loop_exit anchor below it. find_loop_exit/1 ports the interpreter's marker-scanning shape for :break.

Executor bridges

Three new metamethod-coupled bridges plus one error-attribution helper:

• dispatcher_index_method_target/5 — :self resolution via index_value/6
• dispatcher_call_value/4 — generic-for iterator step via call_value/5
• dispatcher_concat/5 — :concatenate slow path with __concat metamethod
• dispatcher_call_function/5 — takes name_hint so nil/non-callable error wording matches :call opcode

call_stack push/pop now happens at every dispatcher call site (both compiled-closure and lua_closure paths) so error-context tests see the same stack the interpreter would.

Verification

• ✅ mix test: 2008 passing (1902 + 51 properties + 55 doctests), 24 skipped, 0 failures
• ✅ mix test --only lua53: 12 passing, 17 skipped
• ✅ mix test test/lua/vm/leak_regression_test.exs: 3 passing
• ✅ closures.exs, oop.exs, string_ops.exs all compile end-to-end (no sub-prototype fallback)

Perf

Hard floor (no workload regresses by >10%): met.

Workload	Dispatcher vs Interpreter
fib(25)	1.15x (vs B5a-v2's 1.17x baseline)
closures(100)	1.22x
oop(100)	1.07x

The ~2% loss on fib(25) vs B5a-v2 is the call_stack push/pop added for error-context parity. The soft target on closures.exs (≥1.5x) is brushed at 1.22x — profile attribution points to closure-construction allocation and upvalue cells, which are post-B5 mutable-storage work.

Test plan

• mix test passes (2008/2008)
• mix test --only lua53 passes
• Leak regression test passes
• closures.exs / oop.exs / string_ops.exs compile end-to-end with no fallbacks
• mix format clean
• mix compile --warnings-as-errors clean