Fix the debugging searches that pruned work they never checked

[cymbalrush]

Search narrowing (search_strategy, validator, comparator):

• _depth_range was computed and logged but never consulted when choosing what to yield, so bisection visited the whole graph at one model execution per side per batch. Both _group_unchecked_nodes_by_depth and _get_next_batch_from_current_level now filter on it.
• The lower bound could advance past unchecked siblings when a level was larger than batch_size; the guard now tests <= the passed depth.
• Status.EXCLUDED was projected to UNKNOWN, which strategies narrow on, so a graph opening with placeholders and views narrowed to depth 0 before a single value was compared. Adds ValidationResult.SKIPPED, which is neither evidence of a fault nor evidence against one.

FX graph diff (graph_diff, graph_match):

• Only top-level fx.Node args were followed, so aten.cat/aten.stack got no incoming edges and their inputs looked unused. Arguments are now flattened through lists, tuples and dicts, and kwargs are wired by name.
• An FX node's identity was its op and target alone -- _attr_digest reads an ir_object an FX node does not have -- so cat(dim=0) and cat(dim=1) hashed identically and diffed as isomorphic. Constant args and the exported result type now feed the digest.

Crashes:

• ComputeDevice.__missing__ is the dict hook, never called by Enum, so an unrecognised residency raised instead of degrading to UNKNOWN.
• _body_counts recorded a call for a callee with no coreai.graph, and _histogram then raised KeyError on the whole histogram.
• Benchmarker: interval id 0 was both the "not timed" sentinel and a real id; the sentinel now sits outside the counter's range and is rejected before lookup. _state transitions moved under the callback lock.