Surface the real error when magic-mode flashing fails

[eliasbakken]

Summary

Two bugs combined to make every magic-mode failure look like a generic "unknown error":

1. flash-from-url ran the download/decompress pipeline as (wget | pv | xz) 2> /tmp/recore-flash-progress - redirecting the whole pipeline's stderr into the progress file swallowed wget's and xz's real error output along with it.
2. Even what little did get logged never reached the UI - goMagic() only ever set state.Error to a generic "An error was encountered during magic. Check log for details" message, which regular users can't act on.

Fixed by capturing each pipeline stage's stderr separately (per-command redirect, not for the whole subshell) and using PIPESTATUS to report whichever stage actually failed, and by having goMagic() surface that specific message to the UI instead of the generic one.

Also dropped wget's -q flag, which turned out to suppress its error output along with its progress noise.

Testing

• make test-bats - un-skipped and fixed the existing (pending) test for this issue; added a stub_stderr test helper since the real fix needed to distinguish stdout (test's canned fake data) from stderr (where wget's real diagnostics go).
• go test ./... - no regressions.
• Live-tested on real hardware via the actual /api/start_magic call:
  • 404 (real file missing) → "Download failed: ... ERROR 404: Not Found."
  • unresolvable host (no internet/DNS) → "Download failed: wget: unable to resolve host address '...'"

Closes #59