perf: fold record-literal projections in the IR optimizer - #157

Closes #153.

Adds reduceObjectProp to the IR rule chain: { foo: 1, bar: 2 }.foo folds to 1, and a projection through a record update takes the patched value, or reaches into the updated record when the field is not patched. At this altitude the fold needs none of the guards the Lua-AST version carries. PropName keys are static labels with no computed-key form, and the type checker rejects duplicate labels, so a plain lookup is exact. A miss is only possible on ill-typed input, where the rule declines.

The payoff of running inside the optimize+dce fixpoint is visible in the goldens. The folded member feeds beta reduction and inlining, so shapes like (function(dictBind) return M.Control_Bind_bind(dictBind) end)(M.Effect_bindEffect) collapse to M.Control_Bind_bind(M.Effect_bindEffect), and dictionary records that only existed to be projected (Data_Either_functorEither, Data_Either_applyEither in the LongEitherBind golden, for example) drop out of the output entirely. The Lua goldens shrink by 205 lines net. The IR golden line counts grow, but that is per-site inlining of single members where the previous pipeline duplicated the whole dictionary table per site and folded it later on the Lua AST.

One deviation from the issue: it proposed deleting reduceTableDefinitionAccessor as superseded, but the goldens show a live trigger the IR rule cannot see. A projection out of a foreign module (ObjectProp (ForeignImport ...), e.g. sub = ({ intSub = ... }).intSub from Data.Ring) only materializes a table constructor during lowering, since the member value is raw Lua source. Deleting the rule regressed those sites, so it stays, with its haddock updated to state the division of labour.

Testing: six new unit tests for the rule (literal, absent field, one-pass composition with sibling rules, patched and unpatched update projections) plus an end-to-end test through optimizedUberModule covering the inliner-created redex. The full suite is green over two consecutive runs after accepting the structural goldens, the IR Optimizer spec was stressed over ten random seeds, and every eval/golden.txt oracle still matches byte for byte (they were never regenerated).