refactor(simd): Phase 4 — extract scalar backend to simd_scalar.rs

Phase 4 of the integration plan in `.claude/knowledge/
simd-dispatch-architecture.md`.

simd.rs shrinks from 1956 → 685 LoC by lifting the inline 1271-line
`pub(crate) mod scalar { ... }` block (impl_float_type! / impl_int_type!
macros + 11 type instantiations + 30+ extra impl blocks for U8x64 /
I8x*/I16x* / U16x32 / I32x16 / I64x8 / F32x16 / F64x8 / U32x16 / U64x8)
out into its own file `src/simd_scalar.rs`. Declared from `simd.rs`
with `#[path = "simd_scalar.rs"] pub(crate) mod scalar;` so the
internal module name stays `scalar` — every `pub use scalar::{...}`
re-export in `simd.rs` continues to resolve unchanged.

Behavior is byte-identical. The file move is mechanical: zero semantic
edits, just outdented one level and wrapped with a module-level doc
comment explaining its role as the non-x86_64 fallback backend.

Why now: the dispatcher should read as a re-export catalog with cfg
arms, not 1.6k LoC of macro expansions. Future tweaks to the scalar
backend (TD-SIMD-5 in the architecture doc) now live in a file named
for what they are.

Author: claude