add fill_tape_padded to skip the internal input copy

[ranflarion]

First half of #469: the padded entry point, without the UTF-8 validation skip (independent, can follow separately).

fill_tape_padded(s, len, buffers, tape) plus INPUT_PADDING. Callers that control the input layout supply the SIMD over-read padding themselves, skipping the O(len) copy fill_tape makes into its internal buffer: +2% to +7% on the shapes measured in #469, growing with document size.

With that copy gone stage 2 reads the same buffer string unescaping writes into, so the read side can no longer be a &[u8] argument (argument protectors under both borrow models, and noalias readonly). It becomes InputView, a pub(crate) pointer+len threaded through build_tape and the per-ISA parse_str, with build_tape taking *mut u8 for the write side. fill_tape still passes a disjoint copy, so that path is unchanged.

One correction to the issue: since 011d699 dropped NUL from the structural-or-whitespace tables the padding is not arbitrary, s[len] has to be structural or whitespace (b' ' works). Documented, and root_scalars_terminated_by_padding covers it.

tests/fill_tape_padded.rs pins the result against fill_tape node-for-node over lane-boundary escape positions, escapes landing against the padding, shared multi-row scratches, error paths and a random corpus. Miri-clean under Stacked and Tree Borrows on both aarch64 and x86_64, which may be useful for #446/#450; full suite, feature matrix, clippy and fmt green.

Happy to rename things or keep InputView out of parse_str's signature if you'd prefer.