`String`, `u32`, etc. as generic type arguments. Builds on the existing
`IntoJsGeneric` / `JsCanon` machinery — adds the inverse direction and
a handful of leaf impls.

* Add `FromJsGeneric` next to `IntoJsGeneric`.
* Add `ref_to_js` to `IntoJsGeneric` for root-level `&T` extern args.
* Weaken `IntoJsGeneric::JsCanon` bound to `ErasableGeneric` so container
  shapes (`Option<T::JsCanon>`, `Vec<T::JsCanon>`) typecheck.
* Add `IntoJsGeneric<JsCanon = Self::Canon>` + `FromJsGeneric` + `'static`
  to `JsGeneric` as supertraits so existing `T: JsGeneric` bounds keep
  working unchanged; the doubled `Canon` associated type propagates the
  canon-erasure invariant to use sites without explicit `where` clauses.
* Generic owned-arg/return codegen routes through `to_js` / `from_js`;
  root-level `&T` routes through `ref_to_js`. For identity impls this
  inlines to the same transmute as before — wasm output for existing
  `JsGeneric` callers is unchanged.
* Leaf impls for `String` / `&str` / numeric primitives in the lossless
  f64 range / `bool`, with `JsCanon = JsValue`. `()` canons to `Undefined`
  so closure return positions participate.
* `Option<T>`, `Vec<T>`, `Box<[T]>` distribute the new traits element-wise.

Reference (`&T` / `&mut T`) paths beyond root-level args, and BigInt-range
integer types (`i64` / `u64` / `i128` / `u128` / `isize` / `usize`), are
out of scope here. Refining leaf `JsCanon` to `JsString` / `Number` /
`Boolean` is blocked by the orphan rule today and waits on a future
js-sys-into-core consolidation; a note documents this.

Tests: `tests/wasm/generics.{rs,js}` gains a small `Cell` fixture with
generic getters/setters exercised at `T = String`, `u32`, `i32`, `f64`,
`bool`, plus a `set_value_ref<T>(&T)` case covering the root-`&T` path.
Guide updated at `guide/src/reference/working-with-generics.md` with
the three-trait split and roles table. `cargo build --workspace` clean;
existing UI test baseline unchanged.