media-export: stream input and output instead of whole-file Blob + BufferTarget (remaining #405 baseline)

[maboa]

Follow-up to #405. The time-stretch no longer buffers the decoded timeline (stream-stretch.js), but the export baseline still holds the whole source and the whole encoded output in memory regardless of rate:

• makeInput fetches the entire source into a Blob before mediabunny reads it (~4–5 GB for a 2 h 1080p file).
• BufferTarget accumulates the full encoded output in one ArrayBuffer, and new Blob([output.target.buffer]) then duplicates it (~2.3 GB + ~2.3 GB transient for the same export).

Proposed direction

• Output: offer mediabunny's StreamTarget backed by the File System Access API (showSaveFilePicker) when available, falling back to BufferTarget + download elsewhere (Safari/Firefox). This is a UX change — the save location is chosen at export start rather than on download — which is why it was split out of media-export: time-stretch path buffers the entire decoded PCM timeline in RAM — multi-GB, OOM on long media #405 rather than bundled.
• Input: for http(s) sources, hand mediabunny a UrlSource/streaming source instead of pre-fetching a Blob; blob:/data: URLs (local files restored from IndexedDB) can stay as they are, since the bytes already live in memory.

Together these drop peak memory for a long 1× export from ~3× file size to roughly the demuxer/encoder working set.