fix(cli): preserve speaker labels in WebVTT cue text

[user-github-me]

Importing a valid WebVTT cue containing ALICE: Hello there silently drops that cue. The parser removes uppercase colon-prefixed lines throughout the file as metadata, so it also truncates multiline captions and changes later cue IDs.

Remove the global metadata-line deletion. Header metadata already sits outside timed cue payloads and is ignored by the existing block parser. Add speaker-label and multiline regressions with an X-TIMESTAMP-MAP header, plus SRT/VTT round trips, preserving cue text, timing, and IDs.

Validation:

• Five new regressions fail before the fix; 220 caption/transcription tests pass afterward, with one existing platform skip.
• Actual built CLI reproduces the missing cue before the fix. Afterward, six imports preserve both cues; twelve repeated SRT/VTT exports remain stable and leave the source bytes unchanged.
• Chrome native TextTrack parsing confirms expected text and timestamps for all twelve original/exported WebVTT cases.
• Rebased on current main 188475aaf; all package builds and the 220 related tests passed again. The rebuilt CLI reproduction, CLI typecheck, repository lint, formatting, comment checks, and all commit hooks passed on that base.

[jrusso1020]

The uppercase-colon metadata strip runs over the whole file, so cue lines like ALICE: Hello are deleted and later cue ids shift. Dropping it is safe: header lines either sit in a block with no timing line or come before the --> line, and only lines after the timing line become cue text. The tests fail on main and pass here. This conflicts textually with #5223 on the same line, so whichever lands second should keep both the CRLF normalization and this removal.

— Rames