refactor(analysis): make analysis visitors linear

[zhen8838]

Why

After #126 merged, dedicated CI grew from 325.64s to 621.75s while adding only four tests. Two analysis paths had become quadratic at model scale: every Call searched the complete Scope tree for its owner, and variance/access relations repeatedly rebuilt shared expression DAG work. This PR restores one-build ownership and traversal while preserving reports byte-for-byte.

What

1. Build Scope ownership and variance tables during the existing normalized walk, using identity-safe integer keys.
2. Convert compute-cost, memory, roofline, and performance collection to family-owned expression visitors without changing timeline emission order.
3. Replace the generic validation dispatcher with a performance checker visitor and keep execution-domain validation before analysis.
4. Fold movement into memory and remove dead structural-occurrence and placement paths.
5. Store each Op-declared access relation on its Scope, project it for each reader's view, and use boundedness instead of exact cardinality where only countability is required.

The deterministic hotspot ledger is: count_val 8,818 -> 3,944; _held_countable 3,986 -> 0; analysis relation rebuilding 1,160 -> 202, comprising 190 primary builds and 12 documented cross-scope fallbacks.

Contract

relations_of now returns the Op-declared, level-independent AccessRelations; local_relations_of returns those relations projected into the caller's type view. This is an intentional semantic name swap across all callers. Analyzer.input_checker is replaced by Analyzer.get_checker(); analyzers without a checker return None. Analysis metadata and rendered reports remain unchanged.

Risk

Scope.one_pass still accounts for 3,056 count_val calls and is intentionally outside this plan. Loop-footprint fallback labels and ordering remain process-identity based; their cross-process nondeterminism is tracked separately in #133 and is not changed here. Validation includes four raw report comparisons, 925 passed plus 4 skipped in the full suite, 156 installed-wheel tests, and identical JUnit test-name sets against the pre-refactor baseline. Dedicated CI passed in 456.00s, down 26.7% from #126's 621.75s but still 40.0% above main's 325.64s; the remaining gap is not hidden by the local host's unrelated compiler load.