Slice Casting

Prefer array chunks plus load/store for ordinary image rows. They accept normal scalar alignment and make tails explicit, as in zenfilters gain.

cast_slice(token, slice) and cast_slice_mut instead borrow a slice of vector objects. They return None when the scalar count or alignment is incompatible. Alignment can differ by backend, so a cast that works for one tier can fail for another. Do not unwrap merely because the input length is a multiple of lanes.

This is a reference-only API exercise; the reviewed zen image loops mostly use array chunks and value loads rather than vector-reference slice casts.

use archmage::prelude::*;
#[magetypes(define(f32x8), v3, neon, wasm128, scalar)]
fn roundtrip_impl(token: Token) -> [f32; 8] {
    let v = f32x8::splat_t(token, 1.0);
    let bytes = *v.as_bytes();
    f32x8::from_bytes_t(token, &bytes).to_array()
}
pub fn roundtrip() -> [f32; 8] {
    incant!(roundtrip_impl(), [v3, neon, wasm128, scalar])
}
assert_eq!(roundtrip(), [1.0; 8]);

as_bytes / as_bytes_mut borrow the native representation; from_bytes and from_bytes_owned construct a value with a token. These are native-endian bit views, not a portable serialized format. Arbitrary float bits include NaNs.

Magetypes vectors do not expose unrestricted Pod/Zeroable construction, which would bypass the token requirement. This does not prohibit using bytemuck for ordinary scalar pixel buffers under that crate's own validity rules.