Deduplicate per-opset-version data in static analyzer rules

Summary

Remove duplicate data across per-opset-version rule ZIP files in the static analyzer. Currently, each opset version ships as a fully independent ZIP, even though the majority of operator rules are identical across consecutive versions. This wastes significant storage and makes rule updates error-prone.

Context

The static analyzer runtime check rules live in src/winml/modelkit/analyze/rules/runtime_check_rules/. Each EP+device+domain+opset combination is a separate ZIP file (e.g., QNNExecutionProvider_NPU_ai.onnx_opset11.zip through opset22.zip). Every ZIP contains 4 JSON files: *_negative_rules.json, *_negative_rules_qdq.json, *_tables.json, and *_tables_qdq.json.

Analysis shows that many operators have byte-for-byte identical rule data across all 12 opset versions (e.g., And, Conv, ConvTranspose, GlobalAveragePool, GlobalMaxPool, InstanceNormalization, LRN). For QNN NPU alone, there are 12 ZIP files totaling ~284 MB, with substantial overlap. The lookup table files (*_tables.json) are even larger and likely have even higher duplication ratios.

Current State

• Rule storage: 20+ ZIP files in src/winml/modelkit/analyze/rules/runtime_check_rules/, one per EP/device/domain/opset combination
• Rule loading: src/winml/modelkit/analyze/core/runtime_checker_query.py:918-946 — loads the ZIP matching the model's exact opset version, no sharing across versions
• Rule generation: src/winml/modelkit/analyze/runtime_checker/result_processor.py:378 — build_op_query_negative_rules_and_table() produces per-operator rules that get bundled per opset
• Data models: src/winml/modelkit/analyze/models/runtime_checks.py
• Example duplication: QNN NPU ai.onnx domain — opset11 ZIP is ~9 MB, opset22 ZIP is ~27 MB, but operators like Conv are identical across all 12 versions

Desired State

A storage format that eliminates redundant operator data across opset versions. Possible approaches (to be evaluated during implementation):

1. Delta/inheritance model: Store a base set of rules and only store per-opset deltas (new operators or changed constraints)
2. Per-operator deduplication: Store each unique operator rule blob once and reference it by hash from each opset version
3. Merged format with version ranges: A single file per EP/device/domain that annotates each operator rule with the opset version range it applies to (e.g., "opset_range": [11, 22])

The loading code in runtime_checker_query.py must be updated to reconstruct the full rule set for a given opset version from the deduplicated format.

Acceptance Criteria

• Duplicate operator data across opset versions is eliminated in the stored rule files
• Rule loading (runtime_checker_query.py) correctly resolves the full rule set for any supported opset version
• No change in analyzer output — existing tests must pass with identical results
• Rule generation pipeline (if applicable) outputs the new deduplicated format
• Storage size of runtime_check_rules/ directory is measurably reduced
• Tests cover the new loading logic, including edge cases (first opset, last opset, opset with unique operators)

Technical Notes

• The _LazyNegRules and LazyDomainTables classes in runtime_checker_query.py use lazy loading — any new format should preserve this behavior for memory efficiency
• The *_tables.json files are the largest contributors to size — prioritize deduplicating these
• Information rules (information_rules/) already use a non-per-opset format and do NOT need changes
• Consider backward compatibility: if the rule generation pipeline is a separate step, the new format needs a migration path from the current per-opset ZIPs
• The com.microsoft domain ZIPs are tiny (~1 KB) and may not need deduplication, but should use the same format for consistency

Related Files

• src/winml/modelkit/analyze/core/runtime_checker_query.py:918-946 — Current per-opset ZIP loading logic
• src/winml/modelkit/analyze/core/runtime_checker_query.py — _LazyNegRules, LazyDomainTables classes
• src/winml/modelkit/analyze/runtime_checker/result_processor.py:378 — Rule generation entry point
• src/winml/modelkit/analyze/models/runtime_checks.py — Rule data models
• src/winml/modelkit/analyze/rules/runtime_check_rules/ — All rule ZIP files
• src/winml/modelkit/analyze/utils/rule_loader.py — Rule loading utilities