Memoize copy-to-output-directory item filtering in graph predictor to improve perf

[derekantrican]

Cache the filtered CopyToOutputDirectory items per dependency ProjectInstance in GetCopyToOutputDirectoryItemsGraphPredictor. When multiple root nodes share transitive dependencies (common in large graphs), the same dependency's items were being filtered repeatedly across 6 item types.

The cache uses ConcurrentDictionary<ProjectInstance, Lazy<CopyItemResult[]>>:

• ConcurrentDictionary for thread-safe access from parallel graph node processing
• Lazy<> ensures the filtering work runs exactly once per dependency under contention
• Results stored as immutable arrays of (InputPath, TargetPath) structs
• Output paths (which depend on the caller's OutDir) are computed at report time

[dfederm]

Do you have any benchmarks to share?

[derekantrican]

@dfederm - I don't right now (been swamped with a lot of other things), but if that's absolutely necessary then I can work on generating some. I would have to figure out what needs to be done to build the library and put it in the right place for the Office build system

[dfederm]

Why change this to an array?

[derekantrican]

With the memoization refactor, BuildCopyItemsCore needs to iterate over each item to build the cached results. Extracting them into a shared array (as opposed to the 6 separate ReportCopyToOutputDirectoryItemsAsInputs calls that existed before) avoids duplicating the list of item names in two places and makes it easier to keep them in sync if a new item type is added later.