Format the tree with ruff, and gate it in CI

[huangziwei]

ruff format over the whole repo — 41 of 46 files. The five already formatted
are the ones somebody had run a formatter over at some point, which is why the
tree currently has two styles in it.

Nothing here changes behaviour, and the diff is large enough that saying so
is not worth much without a check. See "Proof" below.

The settings, and why they are not the defaults

Pinned in pyproject.toml rather than left implicit, so a change to ruff's own
defaults is a deliberate edit here rather than a repo-wide reformat nobody asked
for.

line-length = 88 — ruff's default, and what the already-formatted files
use. 100 and 120 were about the same diff size (2871 and 2611 changed lines
against 2975), so this is not the axis that matters.

skip-magic-trailing-comma = true — this one is a real choice. Ruff, like
black, treats a trailing comma as "keep this exploded". The trailing commas in
this tree are a habit of writing [8, ] and [params[i], ], not a request, and
respecting them produces:

Cinvs = [
    smoothness_kernel(
        [
            params[i],
        ],
        df[i],
    )[1]
    for i in range(len(df))
]

against

Cinvs = [smoothness_kernel([params[i]], df[i])[1] for i in range(len(df))]

isort.split-on-trailing-comma = false comes with it; ruff warns if they
disagree.

extend-exclude = ["*.md"] — ruff formats python blocks inside markdown.
In the README that turns

lnp.add_design_matrix(X_dev, name="stimulus")  # basis will automatically apply to dev set

into a three-line call, because the trailing comment pushes it past 88. Worse
documentation for the sake of a limit the block is not compiled under.

ruff>=0.16,<0.17 in the dev group. It is a CI gate now, and an unpinned
formatter means a release can turn that gate red on a tree nobody touched.

Proof that it changes nothing

Every .py file was parsed before and after and the ASTs compared:

• 36 of 44 produce a byte-identical ast.dump.
• The other 8 have identical structure — comparing ASTs with every string
  constant blanked out gives an exact match — and differ only in whitespace
  inside docstrings, which is the formatter stripping trailing spaces and
  reindenting. For all of them, "".join(old.split()) == "".join(new.split()).

Plus: the suite is 69 passed before and after, and ruff check reports the
same 243 findings both ways, so the formatter neither fixed nor introduced a
single lint issue.

CI

Adds a format job running ruff format --check ..

It does not add ruff check. There are 243 lint findings in the tree, and a
gate that passes because it is not looking is worth nothing. That job goes in
with the change that clears them, which is the next PR — and it is a much less
mechanical one than this: 26 F401 unused imports where some are deliberate
re-exports through import *, 25 unused unpacked variables, 7 bare except:.
Worth keeping separate from a whitespace diff.