chore: add clang-format + codespell lint gate

[balbasty]

Part of #18 — item D6, "No C++ lint/format/codespell". The five Python
repos have run ruff + codespell since they were created; none of the six
C++/CUDA repos had a .clang-format, a .clang-tidy or any spell-check.

This is one of six identical PRs, one per C++ repo (fastfields-kernels,
-cpu-impl, -cuda-impl, -cpu-lib, -cuda-lib, -lib). The three files
added are byte-identical across all six.

What this adds

• .clang-format — derived from the code that is already here, not
  picked off the shelf. I measured the tree before choosing anything:
  4-space is the dominant indent level in every repo (including kernels,
  where a handful of files such as distance/l1.h are 2-space outliers), the
  p90 line length is 68–78 so the column limit is 80, and the code
  consistently spells pointers/references scalar_t * f / DLTensor & out
  (PointerAlignment: Middle). Function braces go on their own line with
  control-flow braces attached (BreakBeforeBraces: Stroustrup), matching the
  existing sources. CUDEV/CUGLOB/CUHOST are declared as AttributeMacros
  so clang-format stops parsing them as a return type and mangling the
  signature that follows, and FF_NAMESPACE_BEGIN/_END as StatementMacros.
• .codespellrc — the same check the Python repos run, with this
  project's vocabulary allow-listed (nd for N-dimensional, numer/denom,
  mone for the literal −1, the t<name> DLTensor locals in the C++ tests,
  unser for M. Unser of the cited B-spline papers, …) and the vendored
  dlpack.h skipped.
• .github/workflows/lint.yaml — a thin, self-contained per-repo workflow
  alongside the existing test.yaml, matching the shape of the Python repos'
  lint.yaml: a blocking codespell job pinned to the same 2.4.3 that
  fastfields/.github's reusable python-lint.yml pins, plus a
  clang-format job. Both carry the concurrency group and
  timeout-minutes the other C++ gates already have.

No source file is reformatted in this PR.

Honest note: the clang-format check is diff-scoped, not whole-tree

The task asked for clang-format --dry-run --Werror over the tracked files,
and asked me to confirm it passes before opening the PR. It does not pass,
and no configuration can make it pass. I measured this rather than assuming
it — with the config in this PR, clang-format would rewrite:

repo	changed lines	files differing
fastfields-kernels	24 727	47 / 54
fastfields-cpu-impl	7 528	12 / 12
fastfields-cuda-impl	9 948	12 / 12
fastfields-cpu-lib	11 517	32 / 32
fastfields-cuda-lib	6 852	18 / 18
fastfields-lib	3 181	20 / 20

Stock LLVM/Google styles are worse (≈89 k changed lines each); the config here
is the best fit I could find (≈63 k). The residual is irreducible: much of the
code is hand column-aligned — DLTensor & out , in the parameter
lists, the backslash-aligned dispatch macro blocks — which clang-format cannot
reproduce by design.

So the whole-tree gate had exactly two possible outcomes, and CONTRIBUTING.md
rules out both: a permanently red main, or a mass reformat bundled into a
tooling PR ("never bundle"). The workflow therefore checks only the lines a
pull request actually touches, via
git clang-format --diff <merge-base> — the incremental-adoption pattern LLVM
itself uses. New and modified code conforms from now on; the one-shot reformat
stays available as its own reviewable change.

Verification

• codespell exits 0 in this repo with the committed .codespellrc
  (checked with the pinned 2.4.3, the same version CI installs).
• The CI clang-format command was run locally against this branch's merge base
  and prints no modified files to format — this PR passes its own gate.
• The gate is not vacuous: appending
  void ff_fmt_probe ( int*x ){if(x){*x=1 ;}} to posdef.cpp and re-running
  the exact CI command produced a failing diff proposing
  void ff_fmt_probe(int * x) with the brace on its own line and int * x
  spacing — i.e. it enforces the project's real style on new lines while
  leaving the untouched legacy lines in the same file alone. (Probe reverted;
  not in the diff.)
• Config parse checked with clang-format 18; the workflow pins
  clang-format-18 so formatting cannot drift with the runner image.

Workstream: claude-jitfields-to-fastfields

🤖 Generated with Claude Code

---

Generated by Claude Code