Fix multicast delivery ordering under concurrent demand

[twittemb]

Concurrent multicast subscribers could advance upstream before the previous result was broadcast, allowing values to arrive out of order or termination to overtake an element. This caused the intermittent CI failure where all three subscribers received [1, 1] instead of [1, 1, 1].

Broadcast each result under the multicast state lock before making the upstream iterator available again. Add regression tests with eight registered subscribers and 100 distinct values, repeated 20 times for both normal completion and failure, and update the changelog. No public API or dependency changes.

Validation:

• Before the fix, the new regressions produced 320 failed subscriber assertions for missing or reordered values.
• After the fix, both CI commands passed locally: build and all 150 tests with code coverage enabled.
• GitHub Actions passed on the identical source tree, including build, tests, coverage generation, and upload.
• git diff --check passed.

Based on the latest main after #59; this PR contains only the multicast fix.