PIPE-3085: hifa_fluxcalflag - cache SPW freq grids and use searchsorted in _get_chanrange
Open the MS once per prepare() call instead of once per line candidate,
and cache the cvelfreqs result per (vis, fieldid, spwid, refframe) via a
new _get_spw_freqs helper. Replace the O(N) linear channel scan with
np.searchsorted for O(log N) boundary lookup. Also convert %-style LOG
calls to lazy-format style and raise ValueError instead of Exception.
