fix(update): make release download resumable to fix slow/stuck updates

The auto-update download used a single 120s *total* request timeout covering
the entire body read, buffered everything in memory, and had no retry or
resume. On a slow link a ~34MB asset could not finish within 120s: the request
was aborted mid-stream, the partial bytes were discarded, and every relaunch
re-downloaded from scratch (issue #293: '34MB update takes 20+ minutes,
never finishes').

Replace it with a connect timeout + a per-attempt budget plus retry using
HTTP Range resume, so a slow-but-progressing download completes across
attempts and a flaky connection recovers from where it left off. Adds an
end-to-end test that drops the connection mid-stream and verifies recovery.