Fixes #46.

Problem

A long straight-line do block desugars to a chain of bind/discard whose continuations nest lexically. Past ~200 levels Lua's parser rejects the chunk with chunk has too many syntax levels (LUAI_MAXCCALLS), so the file fails to load — before any code runs. The prelude test suites hit this in practice.

Approach

A new IR pass Language.PureScript.Backend.IR.MagicDo recognises bind/discard/pure specialised to the Effect or ST monad — whose runtime value is a nullary thunk — and rewrites the nested chain into a flat statement sequence:

function() local x = m1(); local _ = m2(); ...; return last() end

This mirrors the magic-do pass of the upstream JS backend and purs-backend-es, but is realised as a rewrite into existing Let/Abs constructs rather than a new IR node (which would ripple through every RawExp traversal, including the De Bruijn machinery behind #37/#56).

Key points:

• Recognition normalises the application head — resolving module-local aliases, projecting fields out of literal dictionaries, and beta-reducing — until the Effect/ST bind instance is exposed. The optimizer inlines discard to discardUnit.discard = bind, so the chain head is rarely a bare Control.Bind.bind; normalisation sees through that.
• Chunking. A single flat thunk would overflow Lua 5.1's other limit, LUAI_MAXVARS (200 locals/function). The statements are split into nested thunks of ≤150 locals, so both the nesting limit and the local-variable limit stay satisfied.
• Placement. Runs as the final step of optimizedUberModule: after renameShadowedNames (locals are uniquely named, so moving binders into a Let needs no De Bruijn shifting) and after dead-code elimination (so the local _ = statements introduced for discard are not dropped as dead). Folding it into the single pipeline definition means both the compiler and the golden-test harness pick it up — no divergence.
• Effect/ST only. Other monads keep their bind calls (their bind is not "run a thunk"). A long straight-line do in Maybe/Either/State/… can still hit the parser limit; the generic, monad-agnostic fallback is tracked in Generic fallback: flatten deeply-nested do-blocks for non-Effect/ST monads #104.

Testing

• New Golden.LongDoBlock regression — a 300-statement Effect block that previously failed to load now evaluates and prints 1..300.
• All existing eval goldens are unchanged (eval/golden.txt preserved, only golden.ir/golden.lua regenerated), confirming semantics are preserved — including Maybe (MaybeChain) and tailRecM (TailRecM2Shadow), which are deliberately not flattened.
• Full suite green: 237 examples, 0 failures (unit + property + golden + eval + luacheck).

Docs

Adds docs/QUIRKS.md (user-facing Lua-target guide) and notes the residual non-Effect/ST limit there.