…ersistent (#2781). Principle III.

Settings → FlexControl, Settings → USB Cables, and the XVTR overlay button each
called `new RadioSetupDialog(...)` directly, bypassing the m_persistentDialogs
registry that PR #2772 (RadioSetupDialog → PersistentDialog migration) set up.
Symptoms: clicking any two entries opens two windows, frameless toggle skips
the shortcut-opened dialog, and "first opener wins" signal wiring diverges.

This pass routes all four entry points (primary Radio Setup, FlexControl, USB
Cables, XVTR overlay) through showOrRaisePersistent + selectTab, and factors
the txBandSettingsRequested / serialSettingsChanged / sliceLetterDisplayModeChanged
/ QDialog::finished wiring into a single private helper wireRadioSetupDialogSignals
so a fresh construction picks up the full handler set regardless of which menu
clicked first.  The dialog now lives in m_persistentDialogs unconditionally, so
the special-case qobject_cast in setFramelessWindow is removed — the
persistentDialogs loop below already handles it.

Also re-asserts the primary Radio Setup entry's #2772 migration that #2786
silently reverted during a panadapter-layout merge.

Blast radius (codegraph impact on MainWindow::buildMenuBar, max_depth=3):
risk_score=0.7297, 18 high_risk affected (top: MainWindow::MainWindow,
AppSettings::instance, AppSettings::value).  The score reflects buildMenuBar's
breadth across the menu surface; the actual edit is four lambda bodies plus a
helper, none of which alter buildMenuBar's caller-facing contract.  No new
test-target link dependencies (wireRadioSetupDialogSignals is private to
MainWindow.cpp, RadioSetupDialog.h was already #included).

## Summary

Fix spectrum display jumps around dBm range changes.

- Rebase a few incoming FFT frames after dBm scale release so stale
radio-encoded frames do not briefly snap the trace back to the old
display location.
- Apply the same protection to vertical span / scale adjustments, not
just moving the top dBm reference.
- Split manual dBm-range echo handling from auto FFT floor echo handling
so auto-floor can keep its local smooth animation running while waiting
for the radio echo.
- Avoid applying an older auto-floor echo if the local animation has
already moved past it; reissue the current local range instead.

## Cause

The likely regression came from the later dBm stabilization work in PR
#2786 / #2780. That path made the stream decoder switch to the requested
dBm range immediately while the radio could still send several FFT
frames encoded against the old range. It also caused auto-floor movement
to pause behind the pending echo guard, which made the smooth motion
appear as steps.

## Test Plan

- [x] git diff --check
- [x] cmake --build build --parallel 8
- [ ] Manual: drag dBm reference line and release; confirm no snap-back
- [ ] Manual: Ctrl-drag vertical dBm span/scale and release; confirm no
snap-back
- [ ] Manual: enable/open panadapter with auto FFT floor adjustment;
confirm movement is smooth instead of stepping

Build passes. Existing warnings remain unrelated.