http_client: Fully download archives before extracting them - #62653

Installing an external agent (or a language server) from an archive that ships without a SHA-256 checksum used a fragile code path: Zed unpacked the archive directly from the network stream, without ever saving the complete file to disk. If the connection dropped partway through the download, the unpacker saw the data stop mid-file and failed with an archive-extraction error that looks identical to a corrupt package. Users hit this with the Cursor agent, whose ACP registry entry has no checksum and whose package is a 71 MB download: any network hiccup produced a confusing "Failed to Install" error with no hint that the download was the problem.

This change makes both paths behave the same way: every archive is now fully downloaded to a temporary file first, and only then unpacked. The checksum verification still runs when a checksum is available, exactly as before; when none is available it is simply skipped. This removes no protection that exists today, since the streaming path performed no verification either. A failed download now reports a download error ("saving archive contents into the temporary file for ") instead of a misleading extraction error.

The second commit fixes the reason these failures were undiagnosable: agent connect and session-creation errors were flattened with to_string, which keeps only the outermost message of an error chain. The underlying cause (invalid gzip, unexpected end of file, connection reset) never reached the panel, the log, or telemetry. Errors are now formatted with the full context chain for display, logged with their debug representation, and session errors are downcast to LoadError first so typed variants keep their specific handling.

The trade-offs are small: the archive briefly occupies temporary disk space, and installation finishes slightly later because unpacking waits for the download to complete.

Intentionally out of scope, as follow-ups: deduplicating concurrent installs of the same version, and cleaning up staging directories left behind by cancelled installs.

Testing:

• Added a test covering the new no-checksum path (download completes, archive extracts correctly).
• Added a test simulating a connection dropped mid-download, asserting the error names the download rather than the extraction, and that the staging directory is cleaned up.
• cargo nextest run -p http_client --features github-download: 8 tests, all passing.
• cargo nextest run -p agent_ui load_error: 4 tests, all passing.
• cargo fmt --check and script/clippy pass.

Release Notes:

• Improved installation of external agents and language servers: archives are now fully downloaded before being unpacked, so an interrupted download produces a clear error instead of a confusing extraction failure, and install errors now show the underlying cause.