fix(ci): gate AMX modules behind x86_64 + scope nostd build to -p ndarray + rustfmt

Three additional CI failures surfaced by PR #141 (all pre-existing latent
issues, none caused by the original migration):

(1) s390x-unknown-linux-gnu cross-tests — x86-only inline asm leaks

`src/hpc/amx_matmul.rs` + `src/hpc/bf16_tile_gemm.rs` (its wrapper) +
`src/simd_amx.rs` all use `asm!` with `rcx` / `rax` register names. AMX
is an Intel-only ISA (Sapphire Rapids+); the registers don't exist on
s390x / aarch64 / wasm32 / etc. and the asm parser rejects them at
compile time.

Fix: gate all three module declarations behind `#[cfg(target_arch =
"x86_64")]`. On x86_64 CI runners (most) they compile normally and
runtime gating via `amx_available()` already prevents execution on
CPUs without AMX. On non-x86 targets they're skipped entirely.

External consumer audit clean — only `bf16_tile_gemm` uses `amx_matmul`,
and only `amx_matmul` uses `simd_amx`. No cascade gating needed.

(2) thumbv6m-none-eabi nostd — criterion dev-dep tree leaks into
workspace cross-build

CI runs `cargo rustc --target=thumbv6m-none-eabi --no-default-features
--features portable-atomic-critical-section` from the workspace root.
Without `-p ndarray` scoping, cargo evaluates the whole workspace's dep
graph (including dev-deps from ndarray-rand / serialization-tests /
numeric-tests). The bench migration in PR #140 added `criterion 0.5`
as a dev-dep; criterion transitively pulls `serde_core` (which doesn't
declare `#![no_std]`) and `getrandom` (which has its own
no_std-incompatible paths) into the dep tree.

The library `ndarray` itself builds cleanly on thumbv6m no-default-
features (verified: `cargo check -p ndarray --target=thumbv6m-none-eabi
--no-default-features --features portable-atomic-critical-section` is
clean). The CI command just needs scoping.

Fix: add `-p ndarray` to the cargo rustc invocation in the nostd CI
job so dev-dep evaluation is limited to the library's own deps.

(3) cargo fmt --all --check failures

Bench files migrated to criterion in PR #140 used the workspace's
prior-style "brace on next line" formatting (`fn foo(c: &mut Criterion)
\n{`). Stable rustfmt 1.94.1 (pinned per CLAUDE.md) wants "brace on
same line" (`fn foo(c: &mut Criterion) {`). Plus single-statement
closures inlined.

Fix: run `cargo fmt --all` (no manual changes needed). 13 bench files
+ examples/life.rs touched with mechanical formatting changes only;
no semantic changes.

Verification (local, 2026-05-13):
- `cargo clippy --no-deps` (default) → clean
- `cargo check --no-default-features --features portable-atomic-critical-section --lib` → clean
- `cargo rustc -p ndarray --target=thumbv6m-none-eabi --no-default-features --features portable-atomic-critical-section` → clean (CI command as updated)
- `cargo fmt --all --check` → clean

Pre-existing latent bugs (amx asm registers since the module was added,
criterion dev-dep regression introduced by PR #140 itself); fixes land
on the same PR series that exposed them.

Author: claude

.github/workflows/ci.yaml

@@ -107,7 +107,12 @@ jobs:
       - run: rustup target add ${{ matrix.target }}
       - name: Tests
         run: |
-          cargo rustc "--target=${{ matrix.target }}" --no-default-features --features portable-atomic-critical-section
+          # Scope to `-p ndarray` so workspace dev-deps (criterion → serde_core
+          # → getrandom) don't get evaluated against the no_std target. The
+          # library itself builds cleanly under no_std + portable-atomic-
+          # critical-section; only its dev-dependency tree pulls std-requiring
+          # crates that have no business existing in the nostd build.
+          cargo rustc -p ndarray "--target=${{ matrix.target }}" --no-default-features --features portable-atomic-critical-section
 
   tests:
     runs-on: ubuntu-latest

benches/append.rs

@@ -4,26 +4,20 @@ use ndarray::prelude::*;
 fn select_axis0(c: &mut Criterion) {
     let a = Array::<f32, _>::zeros((256, 256));
     let selectable = vec![0, 1, 2, 0, 1, 3, 0, 4, 16, 32, 128, 147, 149, 220, 221, 255, 221, 0, 1];
-    c.bench_function("select_axis0", |b| {
-        b.iter(|| black_box(&a).select(Axis(0), black_box(&selectable)))
-    });
+    c.bench_function("select_axis0", |b| b.iter(|| black_box(&a).select(Axis(0), black_box(&selectable))));
 }
 
 fn select_axis1(c: &mut Criterion) {
     let a = Array::<f32, _>::zeros((256, 256));
     let selectable = vec![0, 1, 2, 0, 1, 3, 0, 4, 16, 32, 128, 147, 149, 220, 221, 255, 221, 0, 1];
-    c.bench_function("select_axis1", |b| {
-        b.iter(|| black_box(&a).select(Axis(1), black_box(&selectable)))
-    });
+    c.bench_function("select_axis1", |b| b.iter(|| black_box(&a).select(Axis(1), black_box(&selectable))));
 }
 
 fn select_1d(c: &mut Criterion) {
     let a = Array::<f32, _>::zeros(1024);
     let mut selectable = (0..a.len()).step_by(17).collect::<Vec<_>>();
     selectable.extend(selectable.clone().iter().rev());
-    c.bench_function("select_1d", |b| {
-        b.iter(|| black_box(&a).select(Axis(0), black_box(&selectable)))
-    });
+    c.bench_function("select_1d", |b| b.iter(|| black_box(&a).select(Axis(0), black_box(&selectable))));
 }
 
 criterion_group!(benches, select_axis0, select_axis1, select_1d);