refactor: read decode words unaligned, dropping the align_to prefix

[bonega]

What

Replace the align_to::<usize>() word-at-a-time structure in both bulk decoders with chunks_exact(USIZE_SIZE) + usize::from_ne_bytes, reading each word unaligned.

Why

The aligned-read design forced a byte-wise decode of the misaligned prefix. LLVM compiles that ≤7-byte prefix loop as a generic %4-remainder loop plus a 4×-unrolled loop, and its unroll cost model cannot be coaxed into the early-exit ladder it emits for the suffix (unreachable_unchecked, length masking, and a hand-unrolled source ladder were all tried — the last one gets re-rolled). Unaligned word loads are a single mov on x86-64/AArch64, and from_ne_bytes compiles to whatever load the target supports, so the prefix problem can simply be removed instead of fixed.

This also makes the len < USIZE_SIZE scalar fallback and the align_of guard unnecessary, and drops the prefix-offset term from the incomplete decoder's error positions.

Results

• Assembly for a monomorphized decode: −41% (complete, 378 → 222 lines) / −34% (incomplete, 478 → 315 lines); the main-loop word read is a single unaligned mov, the tail keeps its early-exit ladder, and the safe try_into().unwrap() leaves no panic branch.
• Benchmarks (decode_checked, decode_lossy, back-to-back vs master with setarch -R/taskset): neutral within noise.

Verification

• cargo test (144 passed), cargo clippy --all-targets -- -D warnings, cargo fmt --check
• Miri writer_stress under Stacked Borrows, Tree Borrows, and big-endian s390x; Miri lib tests --all-features
• thumbv7em-none-eabi no-std builds (bare + alloc)

https://claude.ai/code/session_01D3JXrJnefe7pWs3kJRvTjz