fix: create the parent directory of --lua-output-file

fix: array-literal pattern binders read 1-based Lua indices

The IR uses per-name De Bruijn indices: a `Ref (Local n) k` selects the
k-th enclosing binder named `n`. Two optimizer passes removed a binder
without lowering the references that had skipped over it, leaving them
pointing one binder too far out -- an unbound local that the Lua backend
rejects with `UnexpectedRefBound`. This aborts `Data.Array` compilation
(`foldRecM`, #56):

  * beta reduction drops the applied lambda;
  * dead-code elimination blanks an unused named binder to `ParamUnused`.

Both now repair the survivors with a new `unshift`, the inverse of
`shift 1`. `shift` is refactored to share its scope-tracking traversal
(`overFreeIndex`) with `unshift` so the two cannot drift apart.

Tests: property and example coverage for shift/unshift/substitute,
including the classic capture-avoiding cases (Types spec); a
well-scopedness invariant over a well-scoped expression generator (this
found the DCE sibling bug) plus deterministic regressions for both
passes (Optimizer spec); and an eval golden `Golden.TailRecM2Shadow`
over the `foldRecM` shape that triggered #56.

fix: lower De Bruijn indices when a binder is removed (#56)

The package set, the prelude/strings forks, and this compiler repo moved
from Unisay/* to the purescript-lua org. Update the README badges and
links, the CLAUDE.md references, the test package-set URL, and a golden
test comment. The template and example repos stay under Unisay, so their
links are left unchanged.

chore: point package-set + fork links at purescript-lua org

Convert treefmt.toml → treefmt.nix and expose `nix fmt` / `formatter`:
cabal-fmt, nixfmt, yamlfmt, and dhall (test/ps/{packages,spago}.dhall).

fourmolu is deliberately left OUT of `nix fmt`: its `--haddock-style` is
global (single/multi/multi-compact) with no per-comment or by-length mode, so
any choice rewrites the source's deliberate Haddock mix (short `-- |`, long
`{- | … -}`). Haskell stays a manual `fourmolu -i lib/ exe/ test/` step.

Excludes test/ps/output goldens and .hlint.yaml (yamlfmt strips its trailing
template comments). CI gains a content-based format check
(`nix fmt && git diff --exit-code`). The pre-commit hook is now a tracked
.githooks/pre-commit wired via `git config core.hooksPath` — worktree/
submodule-safe, never clobbers an existing .git/hooks, and surfaces `nix fmt`
failures instead of swallowing them. cabal-fmt/nixfmt reflow
pslua.cabal/shell.nix.

The format check ran in the tests job after 'cabal test all', which
regenerates test/ps/output/*/corefn.json (treefmt excludes that path).
'git diff --exit-code' then tripped on the regenerated corefn rather than
on any formatting issue. Move it to a dedicated job with a clean checkout.

Points test/ps at the new Lua package-set release (the FFI-audit fixes) and
regenerates the goldens. Churn is cosmetic and behaviour-preserving:

- golden.ir: only `.spago/<pkg>/<version>` source-path bumps.
- golden.lua: only foreign-table content from the bumped forks (chiefly
  effect's pureE/bindE reformatting; plus the audited FFI fixes where they
  survive DCE).
- one corefn.json: a source-comment org rename (Unisay -> purescript-lua).
- All eval/golden.txt are unchanged — runtime output is identical across
  every executable golden, so no behaviour changed.

The golden suite spiked to ~21 GiB (and got OOM-killed) when many goldens
differed: hspec held two full pretty-printed blobs plus their diff per
failure. Two changes:

- Test.Hspec.Golden now reports a bounded, line-oriented summary by default
  (first differing line + a small window + line counts + the actual-file
  path); set PSLUA_GOLDEN_FULL_DIFF=1 for the complete expected/actual diff.
- The test suite's RTS opts go from -N (all 24 cores → parallel-GC RSS scales
  with core count) to -N4 -c -M16g, so peak residency is bounded and a
  runaway heap-overflows cleanly instead of OOM-killing the machine.

Full suite: 233 examples, 0 failures.