Fix Windows test failures from PR #201

[prabhu]

Follow-up to #201 (merged before its Windows legs finished, since only pr-tests (ubuntu-latest, 25) was a required check at the time). Two fixes:

jssrc2cpg: expect / project-relative names in the $lib import tests on every OS

SvelteKitAliasImportTests still built its expectations with File.separator, but #101's commits a255183c and 55d895e4 made method full names carry / project-relative names on every OS, so the two resolution tests failed on every Windows job:

"src/lib/api.js::program:getArticle" did not include substring "src\lib\api.js"

js tags: scope JsIndexes' assignment index by graph

A node's equals/hashCode cover its id only, never its graph, and every CPG numbers its nodes from the same range, so the static WeakHashMap[Method, ...] matched methods across CPGs built in the same JVM. One test's assignment map then served another's lookup:

• with the foreign graph still open, a local resolved to the previous test's expression and EasyTagsPass tagged xhr.send and seen.delete as SDK actions — the intermittent JsDomRecognizerTest failure seen on CI (Windows/JVM 26 on the PR build, macOS/JVM 23 on the merge-group build);
• with the graph closed, accesses threw MVStoreException: This store is closed.

The index is now keyed by graph first (as methodIndexes already was), making the inner by-node key exact since ids are unique within a graph. Verified with a 30-round close/reopen stress loop (previously failed from round 2) plus full suites: x2cpg 136 and jssrc2cpg 679 tests green.

Branch protection on main now requires one pr-tests leg per OS (ubuntu/windows/macos, JVM 25), so the queue can no longer merge while another OS is failing.