Previously, Zinc could produce nondeterministic analysis output because binaryClassName used put(binary, className) from concurrent externalLibraryDependency callbacks. The final value depended on callback scheduling, so binaryClassName stores one representative class per binary JAR in a concurrent map, but representative selection was non-deterministic.

This PR replaces put with deterministic merge using updateWith, choosing the lexicographically smallest class name for each binary. libraryClassName only needs a stable representative class per binary for lookup/stamp checks; selecting a deterministic representative preserves behavior while removing scheduler dependence. This removes one source of nondeterminism in Zinc analysis serialization and downstream build cache hashes, which was blocking attempts at reproducible builds in Mill (e.g. com-lihaoyi/mill#4642)