feat: add crates/transcript with TranscriptAccumulator

Summary

Extracts the TranscriptAccumulator into a new standalone crates/transcript shared crate — the crate-only portion of #4064, with no plugin or desktop app integration.

The accumulator processes streaming ASR StreamResponses into clean, deduplicated transcript words using a two-level design:

• Within a response: assemble aligns tokens to the transcript string (sole oracle for word boundaries — no timing heuristics)
• Across responses: stitch uses a 300ms timing heuristic to merge words split across final responses (e.g. Korean particles)

Per-channel state handles watermark-based dedup, partial word splicing, and held-word stitching. 37 unit tests pass, including fixture replay tests against real Deepgram and Soniox data (English + Korean).

Nothing depends on this crate yet — integration into plugin-listener/plugin-listener2 is a follow-up.

Review & Testing Checklist for Human

• Verify spacing_from_transcript greedy-forward find handles repeated substrings in transcripts correctly (e.g. "the the") — mismatch could silently assign wrong spacing
• Review the 300ms threshold in should_stitch — this is the heuristic for merging words split across responses; confirm it works for your provider mix
• Confirm splice boundary conditions (<= / >=) don't drop or duplicate words at exact timestamp boundaries
• Run cargo test -p transcript locally to confirm fixture tests pass against current hypr-data

Notes

• Files were extracted directly from the refactor/transcript-accumulator-to-shared-crate branch (PR refactor: move TranscriptAccumulator to shared crate, migrate batch streaming path #4064) via git checkout
• Uses edition = "2024" — matches the convention on that branch
• Link to Devin run
• Requested by @yujonglee