lowering: a mut assigned inside a CE block gets its global/nonlocal (#122)

[simontreanor]

A built-in CE body (async/seq/result/option) lowers to its own nested def/async def, but the closure-capture declarations (DESIGN.md §3) were only computed for function bodies. So flag <- true inside an async { } block bound a silent local in the coroutine and the module binding never changed — the quiet miscompile #122 reports, where a finished host idled forever on a flag that could not flip. The result { } read-modify-write form (attempts <- attempts + 1) hit the louder UnboundLocalError instead; every built-in CE had one form or the other.

The fix computes the same global/nonlocal prelude a function body gets, for all four built-in CE frames:

• scan_ce_scope mirrors scan_scope over CE items — let/let!/for targets are the frame's binders, a for body is the same Python frame so it is entered, nested functions and CEs still are not.
• The split (captured name in an enclosing function frame → nonlocal, else global) is factored out of lower_fn_body into classify_captured/capture_decls and reused; declarations are named after lowering so a renamed mut slot is declared under the name the closure assigns, as before.
• The CE's own binders are pushed on fn_local_stack while its items lower, so a lambda nested inside the block classifies its captures against the CE frame too (and the fold/decode inlining guards, which union the stack as an over-approximation, only get more conservative).

The issue's repro now emits global flag inside the async def and prints True. User builders were already covered: they desugar to lambdas, whose statement-bodied form goes through lower_fn_body.

Tests: emitted-shape assertions for the global (async, result) and nonlocal (enclosing-function mut) cases, plus e2e runs of the filed repro and a seq counter. DESIGN.md §3 records that a CE body is a frame under the same rule; ROADMAP.md gets the feedback item.

Closes #122