Unbox non-escaping Ref/STRef cells to mutable Lua locals - #315

Closes #239.

A Ref/STRef cell compiles to a one-field heap table: the foreign new allocates {value = v}, and every read/write/modify pays a table field access. This PR adds a codegen-time pass (Language.PureScript.Backend.Lua.RefUnbox) that lowers a non-escaping cell to a plain mutable Lua local: new v becomes local r = v, a read run becomes r itself, and write/modify runs become assignments. The load-bearing observation is that a Lua local captured by inner closures is an upvalue — a shared, mutable heap slot with exactly the aliasing behaviour of the {value = …} table — so occurrences under lambdas (loop bodies, nested magic-do chunks) need no special treatment, and the only thing the table provides that a local cannot is first-class identity. The analysis therefore asks one question per cell: is every occurrence the reference argument of a recognised operation? Any other occurrence (stored, returned, passed to an unknown function) keeps the cell boxed, so a miss is a missed optimisation, never a miscompile.

What lowers. Recognition is by qualified name — the identity that survives linking, the same scheme magic-do and the native-loop matcher use — over the cell primitives of Effect.Ref and Control.Monad.ST.Internal (_new/new, read, write, modifyImpl) plus the ST functor's foreign map_ as a wrapper, because void (STRef.modify f r) (the ubiquitous modify_ shape) inlines to map_ (\_ → unit) (modifyImpl f r). Only run positions lower (the application carries magic-do's EffectRunArg marker, or is the body of a literal loop lambda that NativeLoop.runStatements runs per iteration); an operation thunk used as a first-class value keeps the cell boxed. A literal modify function whose body ends in a manifest {state, value} record is beta-reduced at emission — parameter bound to the local, Let spine emitted as statements, state field assigned, value field passed on — so the per-iteration record is never allocated either.

Before/after, the ST accumulator loop from Golden.NativeLoopsST.Test (quoted verbatim from the committed goldens):

local acc = Control_Monad_ST_Internal_new(0)()
for i = 0, n + 1 - 1 do
  Control_Monad_ST_Internal_modifyImpl(function(s_S_0)
    local sPrime_S_0 = s_S_0 + i
    return { state = sPrime_S_0, value = sPrime_S_0 }
  end)(acc)()
end
return Control_Monad_ST_Internal_read(acc)()

local acc = 0
for i = 0, n + 1 - 1 do
  local s_S_0 = acc
  local sPrime_S_0 = s_S_0 + i
  acc = sPrime_S_0
end
return acc

Soundness guard. The new Golden.RefUnbox.Test pins a cell stored inside another cell (outer <- STRef.new inner): the stored inner keeps its {value = …} table while outer still unboxes, and boxed operations on an aliased cell keep going through the foreign implementation. Golden.NativeLoopsST.Test's countDown pins the per-cell granularity of the value-position restriction: its steps cell (used only in run positions) unboxes while its value cell (read under map_ inside the while condition — a first-class thunk) stays boxed. All eval/golden.txt oracles — which are never auto-accepted — are byte-identical, so runtime behaviour is unchanged across the whole golden corpus, including the write-returns-written-value ST semantics and Effect write's nil result.

Measurement (issue's verification asks for the allocation-counter drop). A new macro bench Bench.RefLoop — a hot ST loop accumulating through a local STRef — with committed counter oracles: the static table-allocation census (tnew_census) drops from 2 steady-state sites to 1 (the survivor is the dead foreign new closure body, never called), and the trace report shows the hot loop trace-compiled (JFORL) instead of aborted on per-iteration closures. Wall-clock at n=100000 on this machine: PUC Lua 5.1 0.0347s → 0.0009s, LuaJIT 0.024s → 0.0001s (medians of bench/tools/run_macro.lua). All pre-existing counter oracles are byte-identical, i.e. no other benchmark's shape moved.

Mechanics. fromIR carries a new set of unboxed cell names: the Let case decides unboxability per allocation binding (scanning the remaining bindings and body), lowers operation-run statements in place (scoped in do … end blocks whenever the lowering declares locals, keeping magic-do's ~150-statement chunks under Lua's 200-active-locals cap), and threads the extended set through everything downstream; an expression-position run compiles to a chunk ending in a return of the run's value. Two NativeLoop adjustments: runStatements now statementizes chunk-lowered runs (an unboxed-cell run returns its value in a tail return, which a run-for-effect must rewrite into evaluation statements — previously only return-free loop chunks reached that path), and dropValue additionally drops a discarded field read off the module-scope table (built by codegen as a plain table, so no metamethod can fire), which the void wrapper's unit value otherwise leaves behind as a dead local per loop iteration.

The two pre-existing -Wname-shadowing warnings in Lua.hs (both \group lambdas shadowing Relude's group) surfaced during the rebuild and are fixed by renaming to dispatchGroup, per the repo's no-tolerated-warnings rule.