refactor(check): layer check into selection / inputs / execution / comparison

[zhen8838]

Why

• check eagerly materialized every declared weight before selecting a leaf, so checking one runnable function could allocate the union of unrelated sibling weights and fail before execution.
• Selection, input acquisition, execution, and comparison were coupled behind overlapping flags and executable IR wrapper objects.

What

• Split check into explicit selection, activation/weight acquisition, concrete execution, and report comparison layers.
• Added lazy runtime resources with weak safetensors read caching and strongly memoized deterministic random draws shared across subtree views.
• Moved semantic execution into the evaluator through a transient Reading, preserving lazy child resources and host-side orchestration.
• Restored reports for input provenance, declared and actual dtypes, selected variants, pinned dimensions, reference norms, warnings, and matching text/JSON facts.

Contract

• Replace --input / --ckpt with independent --inputs random|files:A,B and --weights random|ckpt:DIR axes.
• LoadedModule is frozen data containing module, resource, and child modules; it neither owns materialized tensors nor executes functions.
• Module.<fn> resolves to the Function / PrimFunction IR node itself, as specified by core-ir section 1.1.
• Remove Module.forward and Module.__call__; semantic execution goes through the evaluator only.
• evaluate accepts a LoadedModule, supports explicit function=, and takes execution placement from caller-provided device=.
• Export evaluator.reading(loaded, device=...) as the public host-orchestration view.

Risk

• Ten files under examples/ still use the old executable-loaded-module API. They are outside pytest and require a separate migration plan.
• One-to-many load_group aliases and the two-level runtime twin base-class structure are intentionally unchanged.