Summary

Closes #3211.

PR #2972 (cf308d6) added auto-download for ChannelContent::File so
agents receive a saved-on-disk path. The ChannelContent::Voice arm
of dispatch_message was missed — voice notes still flowed through as
raw [User sent a voice message (Xs): <url>] text, forcing every
agent to curl the URL and juggle workspace-allowed paths before
media_transcribe (or any other audio tool) could read the file.

This mirrors the File arm:

• Derive a filename from the URL basename (filename_from_url),
  falling back to voice.ogg when the URL has no usable basename.
• Call the existing download_file_to_blocks with the same
  channels.file_download_max_bytes / channels.file_download_dir
  config knobs and the same disk-cleanup sweeper.
• Prepend a ContentBlock::Text carrying [Voice message (Xs)] plus
  any caption so the model still knows it's voice and any user
  caption survives the URL→path swap.

When the download fails the existing fallback formatter ([User sent a voice message …]) still runs — same shape as the File arm.

Test plan

• cargo test -p librefang-channels --lib — 803/803 pass,
  including 3 new tests for filename_from_url (basic, query/fragment
  stripping, empty/unparseable URLs).
• cargo clippy -p librefang-channels --lib --all-targets -- -D warnings — clean.
• Build: cargo build -p librefang-channels --lib — clean.
• Live: send a Telegram voice note to a channel-bridged agent;
  observe a [File: voice.ogg] saved to /tmp/librefang_uploads/...ogg
  block in the inbound message and that media_transcribe reads it
  directly without an agent-side download. (Tested manually on the
  reporter's setup — works end-to-end with media_transcribe against
  the saved path.)

Notes / follow-ups

• ChannelContent::Video and ChannelContent::Audio likely have the
  same gap — left for a follow-up to keep this PR scoped.
• WhatsApp's outbound send_voice (feat(whatsapp): add send_voice and dm/group message policies #3012) is the symmetric path; this
  PR is purely the inbound side.