feat(meta): re-export the streaming surface from the crate root

[Crauzer]

The bin-streaming reader is merged and unreleased. Before it ships, two gaps in its
public surface are worth closing, because both are compiler-proven rather than a
matter of taste, and neither can be fixed later without a second version bump.

The root under-exported stream

stream publishes twenty-one names; the root re-exported eleven. ValueView was
among them but ContainerView, MapView, StructView and OptionalView — the four
you reach by matching on it — were not, so every real walk imported from two places.

The root now mirrors the module's own pub use list. This is what
section 4
already claimed ("re-exported from the crate root") and what #207 promised, so it
closes a gap rather than opening a question. No collisions: tree, error, file
and data_override share none of the ten new names.

concrete carries what Rust cannot infer

A type parameter's default applies in type position but never in expression position.
That makes concrete::BinStream load-bearing rather than a convenience:

let mut stream = ltk_meta::BinStream::mount(File::open(path)?)?;
// error[E0283]: type annotations needed for `BinStream<File, _>`

LruObjectCache::new is the same case, so it joins it. NoCache is re-exported
rather than aliased, because an associated function resolves through a type alias
but a unit struct's value constructor does not.

The views are deliberately not aliased. They are only ever named in type
position, where the default already applies, so the root re-export above is enough —
a second path would rebuild the two-import-site problem this removes.

One trait was missing its default

ObjectCache was the only type in the module without M = NoMeta. A consumer
holding a policy in a struct field had no inference to lean on and had to write
Box<dyn ObjectCache<NoMeta> + Send>, importing NoMeta from a third path. It now
carries the default the rest of the surface has. Nothing published contains this
trait, so it costs no compatibility.

Guard

tests/import_surface.rs names every streaming type from the root or concrete and
nowhere else, mounts with no turbofish and no annotation, names the four sub-views in
type position, and holds a cache policy in a struct field. It asserts real fixture
content, so it fails on a lost export and on a lost default, not merely on a
rename.

Docs

Section 4 read as though moving M onto the handle resolved the annotation problem.
It does not — mount is expression position — and it promised concrete aliases for
two types that do not exist. Both corrected in place; S12 keeps its ID; ADR-0010 is
untouched and still carries the argument.

#210, #211 and #217 are parked with the evidence against each rather than with
"later". All three stay open.

Verification

cargo fmt --all --check, cargo clippy -p ltk_meta --all-targets -- -D warnings,
cargo test --workspace and cargo doc -p ltk_meta --no-deps (under
RUSTDOCFLAGS="-D warnings") all clean.

Also here, as a separate chore: commit: cargo clippy --workspace --all-targets -- -D warnings
was failing at crates/ltk_anim/src/rig/write.rs:71 on unnecessary_sort_by — a clippy 1.97
regression in a file untouched since 2025-09-26, so main was already red. sort_by_key with
Reverse is the same stable descending sort by the same key, so the bytes written do not move.
It is chore rather than fix on purpose: nothing was broken for a user, and release_commits
would read fix as reason to cut an ltk_anim release this does not warrant. The workspace
clippy gate is now green.