perf(ssr): only record import-binding identifiers in ssrTransform

[shulaoda]

During the AST walk in ssrTransform, every reference identifier was recorded with a full parent-stack snapshot, even though only import-binding identifiers are ever acted upon.

Problem

• For each reference identifier, walk pushed [node, parentStack.slice(0)] into identifiers, then the final BFS pass re-ran isInScope and called onIdentifier.
• onIdentifier immediately returns for any name not in idToImportMap. Since most identifiers in a module are locals/globals, the vast majority of those parentStack.slice(0) allocations and second isInScope scans are wasted.

Fix

• Pass idToImportMap into walk and gate the record with idToImportMap.has(node.name) as the first (cheapest) check. Non-import identifiers are no longer snapshotted or re-scanned. idToImportMap is fully populated before the walk and only read during it, so output is unchanged.

Tests

• No new test is included: this is a behavior-preserving optimization with byte-identical output, so there is nothing new to assert. The existing ssrTransform suite (71 tests) passes unchanged, confirming identical output.
• Measured with a microbenchmark (1 import + ~2000 local references) via vitest bench: the worst-case path improved from ~27.5ms to ~6.5ms mean (~4.3×).