fix(queue): support voice and media prompts

[Marenz]

• queue transcribed voice, photos, supported documents, and media groups while a run is active
• retain at most 5 queued items and 20 MB of aggregate raw Telegram media bytes, measured from file_size before media preparation
• preserve albums as one queue item and keep queue entries individually removable

npm run lint, npm run typecheck, npm run build, and npm test pass.

[grinev]

@Marenz thanks for PR. Overal it looks good, but please fix 2 issues before merging:

1. Major — queued media needs an explicit aggregate size limit.

MAX_QUEUED_PROMPTS = 5 limits only the number of queue items. It does not limit the amount of media kept in memory. A queued item
may contain several files from an album, and every file is stored as a base64 data: URI. Five albums can therefore retain hundreds
of megabytes or more and may terminate the Node.js process with an out-of-memory error.

Please add an aggregate limit for queued media. I recommend starting with 20 MB for all media currently held by the queue, not 20 MB
per queue item. The exact value may be adjusted, but it must be explicit and documented. The limit should be checked before
downloading/preparing files, and a full queue should reject the message before downloading it.

2. Moderate — documentation must be updated in this PR.

README.md and PRODUCT.md still describe the message queue as supporting text messages only. This PR changes user-visible behavior
by adding voice messages, photos, documents, and media groups. Please update the documentation to describe the supported media types,
the queue limit, and the media-size restriction. This is required for the feature and should not be left as an optional follow-up.

[Marenz]

Rebased onto main; the 20 MiB aggregate cap is now enforced from Telegram file_size before any download or data-URI preparation — photos, documents including PDF/office, and albums (rejected as a whole). Unknown file_size is rejected rather than admitted. Accounting is released on dispatch/removal. README/PRODUCT updated.

Voice is unchanged: it's transcribed on arrival and the queue keeps only text, so no media bytes are retained.

Lint, typecheck, build and the full suite (1695 tests) pass.

[grinev]

@Marenz thanks for contribution!!