Reuse normalization scratch buffer

[vicotrbb]

What changed

• reuse one hidden-size normalization scratch vector across attention, feed-forward, and output normalization in a single-stream token step
• preserve the public allocation-returning normalization API while adding an internal refill path
• add an allocation-reuse unit test, changelog entry, and reproducible benchmark note

Why

A single-stream token step previously allocated a fresh normalization vector twice per transformer layer, plus once for the output projection. These stages do not overlap, so one vector can be safely refilled without changing arithmetic order or model output.

Impact

For a model with L layers, normalization scratch allocations per generated token fall from 2L + 1 to 1. Five interleaved 128-token benchmark pairs produced identical complete token traces. The timing median improved by 0.57 percent on a contaminated host, so the timing is documented as diagnostic only.

Validation

• cargo fmt --all --check
• cargo clippy --workspace --all-targets --all-features --locked -- -D warnings
• cargo test --workspace --all-targets --locked
• cargo test --workspace --all-targets --all-features --locked
• RUSTDOCFLAGS="-D warnings" cargo doc --workspace --all-features --no-deps --locked
• cargo test --workspace --all-features --doc --locked
• 68 available release-mode real-model tests across five GGUF artifacts
• Python eval and release-tool tests
• documentation and repository policy checks
• cargo audit, cargo deny, cargo machete, and cargo tree --duplicates
• package verification for ferrite-fixtures and ferrite-model, plus ferrite-inference archive listing