Backend-scoped Like fast path, tighten Dimension constructor, ShapedLike docs

- Scope Like fast-path trust per backend: JAX trusts (ndarray, jax.Array),
  Torch trusts (ndarray, Tensor). Foreign-backend arrays now fall through
  to the slow path where actual convertibility is verified via asarray.
- Add trusted_types parameter to _ArrayLikeChecker, _ArrayLikeFactory,
  and make_array_like_type (default None preserves backward compat).
- Reject non-str/non-int in Dimension.__new__ (float, None, object).
- Reject +Scalar, +__, +~N in Dimension.__pos__.
- Guard bare "+" and "+__" in _dim_spec broadcastable parsing.
- Rename _ValueExpr to Value, remove TYPE_CHECKING branches for dims.
- Tighten broadcastable docs in README and notebook.
- Add ShapedLike runtime-vs-static approximation note to README.

Author: acecchini

README.md

@@ -235,7 +235,7 @@ def last_dim(x: F32[~__, C]) -> F32[~__, C]:
 
 ### Broadcastable dimensions
 
-Apply `+` (unary plus) to any dimension to make it **broadcastable** — size 1 always matches, regardless of the bound value:
+Apply `+` (unary plus) to a named dimension, integer, or expression to make it **broadcastable** — size 1 always matches, regardless of the bound value:
 
 ```python
 from shapix import N, C
@@ -447,6 +447,8 @@ Dtype compatibility uses NumPy's `same_kind` casting rules by default: `int32` c
 
 **Available:** `BoolLike`, `I8Like`, `I16Like`, `I32Like`, `I64Like`, `U8Like`–`U64Like`, `F16Like`, `F32Like`, `F64Like`, `F128Like`, `C64Like`, `C128Like`, `C256Like`, plus category aliases `IntLike`, `FloatLike`, `NumLike`, `ShapedLike`, etc.
 
+> **Note:** `ShapedLike` has the same runtime-vs-static approximation as `Shaped`: runtime accepts any dtype, static alias narrows to bool | numeric.
+
 Like types are also available in JAX and PyTorch backends:
 
 ```python