fix(simd): revert avx2 predicate tightening — env RUSTFLAGS overrides v3 config

The previous commit tightened the x86_64 dispatch arm to
`target_feature = "avx2" + not(avx512f)`. The intent was to make
"x86-64 baseline + AVX2 wrappers" a compile error rather than a SIGILL.

CI green-mode disagreed: `.github/workflows/ci.yaml` sets a global
`RUSTFLAGS="-D warnings"` env that overrides the rustflags from
`.cargo/config.toml` entirely (cargo doesn't merge env + config
rustflags — env wins). So in CI the v3 baseline never takes effect,
x86-64 generic / SSE2 is what builds, `target_feature = "avx2"` is not
set, and the tightened arm leaves no matching dispatch path → consumer
references to `crate::simd::F32x16` fail to compile.

The pre-existing wider `not(avx512f)` predicate works at x86-64
baseline because the inner intrinsics in `simd_avx2.rs` use per-function
`#[target_feature(enable = "avx,avx2,fma")]` annotations — the OPS
gate themselves at the symbol level, struct fields like `__m256` /
`__m256i` are core::arch type declarations that don't require AVX/AVX2
at the type level (only at execution).

Reverting the predicate. The cargo configs added in the previous commit
stay — they're the documented opt-in affordances. Local `cargo build`
without env override gets v3; CI runs at baseline + per-function
target_feature; explicit AVX-512 via `--config .cargo/config-avx512.toml`.

Author: claude

src/simd.rs

@@ -279,16 +279,21 @@ pub use crate::simd_avx512::{f32_to_bf16_batch_rne, f32_to_bf16_scalar_rne};
 #[cfg(all(target_arch = "x86_64", target_feature = "avx512bf16"))]
 pub use crate::simd_avx512::{BF16x16, BF16x8};
 
-// AVX2 baseline arm — selected by the `x86-64-v3` cargo default. Requires
-// `target_feature = "avx2"` explicitly: building x86_64-without-AVX2 (the
-// generic `x86-64` baseline = SSE2) would otherwise pick this arm and
-// then SIGILL on the `__m256` / `__m256i` intrinsics inside the wrappers.
-// Whoever wants no-AVX2 must pick the scalar fallback path (currently
-// non-x86 only — see TD-SIMD-7 in the architecture doc).
-#[cfg(all(target_arch = "x86_64", target_feature = "avx2", not(target_feature = "avx512f")))]
+// AVX2 baseline arm — selected by the `x86-64-v3` cargo default. The
+// predicate is `not(avx512f)` rather than `avx2 + not(avx512f)`: the
+// inner intrinsics in `simd_avx2.rs` use per-function `#[target_feature
+// (enable = "avx,avx2,fma")]` annotations, so the OPERATIONS gate
+// themselves at the symbol level even when the consumer build target
+// is x86-64 baseline. The struct-field types (`__m256` / `__m256i`)
+// are core::arch declarations and don't require AVX/AVX2 at the type
+// level — only execution does. Keeps GitHub CI green (it runs with
+// `RUSTFLAGS="-D warnings"` env, which overrides our v3 config.toml,
+// landing on x86-64 baseline → the previous tighter `avx2` predicate
+// left no matching arm).
+#[cfg(all(target_arch = "x86_64", not(target_feature = "avx512f")))]
 pub use crate::simd_avx512::{f32x8, f64x4, i16x16, i8x32, F32x8, F64x4, I16x16, I8x32};
 
-#[cfg(all(target_arch = "x86_64", target_feature = "avx2", not(target_feature = "avx512f")))]
+#[cfg(all(target_arch = "x86_64", not(target_feature = "avx512f")))]
 pub use crate::simd_avx2::{
     f32x16, f64x8, i16x32, i32x16, i64x8, i8x64, u32x16, u64x8, u8x64, F32Mask16, F32x16, F64Mask8, F64x8, I16x32,
     I32x16, I64x8, I8x64, U16x32, U32x16, U64x8, U8x64,