Instrument centered coordinates for Moon/Sun sidelobe cuts

[chervias]

This change was implemented with Claude Code (Opus 5.5).
It requires the companion so3g PR that adds the det_left mode to
the projection engine (Assembly/ProjEng "det_left").

Summary

Adds a way to build a pointing matrix in instrument-centered
coordinates of a source (Moon, Sun, or any name accepted by
planets.get_source_pos), and uses it to flag samples from per-wafer
sidelobe masks delivered as observation metadata.

Instrument-centered coordinates give the position of the source
relative to each detector, on axes fixed to the telescope structure
(the focal plane xi/eta axes at zero roll). Unlike the previous
object-centered frame (detector pointing relative to the source), a
sidelobe that is fixed to the instrument stays fixed in this frame.

Quaternion chain

For detector d and sample t:

q[d,t] = A[d] * X[t]
X[t] = Rz(roll[t]) * q_bore[t]^-1 * q_src[t]            (sight.Q)
A[d] = Ry(pi/2) * Rz(roll0) * rotation_xieta(xi, eta, 0)^-1
       * Rz(-roll0)                                      (fp.quats)

• q_bore is the standard CelestialSightLine.az_el boresight
  (qpoint * Rz(pi + roll)); q_src = rotation_lonlat(ra, dec) of the
  source. Only q_bore^-1 * q_src enters, so the horizon->celestial
  rotation cancels and the full-precision celestial pointing is used.
• Rz(roll[t]) removes the focal-plane roll so the axes are fixed to the
  telescope structure; conjugating q_det by Rz(roll0) (circular mean
  roll) expresses the detector offsets in the same axes. Exact for
  constant roll; a warning is logged if roll varies by more than 0.1
  deg from its mean.
• Detector quaternions are rebuilt with gamma = 0, since with q_det^-1
  on the left gamma would rotate each detector's map about the
  detector.
• Ry(pi/2) moves the detector from the pole to (lon, lat) = (0, 0).

Map conventions: each detector is at (0, 0); for small offsets
lat ~ eta_src - eta_det and lon ~ -(xi_src - xi_det), so in a standard
CAR geometry right = +az and up = +el (commanded). For LAT observations
past the zenith (el > 90) the axes follow the structure and are
therefore flipped relative to the sky.

Changes

• coords/pmat.py:
  • P has a det_left attribute, passed to Assembly.attach in _get_asm.
  • P.for_tod(..., instrument_centered=) builds the
    pointing above. Requires comps='T', geom given, rot=None and the
    boresight in az/el/roll (boresight_equ is not supported); raises
    ValueError otherwise.
  • New get_instrument_centered(sight, fp, roll, source, timestamps,
    site) returning the (sight, fp) pair for det_left projection.
• coords/planets.py: new get_source_quat(source, timestamps, site,
  dt=600): source RA/Dec from get_source_pos on a grid of at most
  dt seconds, RA unwrapped and linearly interpolated to each sample,
  returned as rotation_lonlat quaternions.
• coords/sidelobes.py: rewritten. Masks are no longer fixed files per
  object; they come from the observation metadata as aman.sidelobes
  (Enmap metadata loader): .map is an enmap [nfields, ny, nx] of 0/1
  in the instrument-centered coordinates of each object, labelled by
  the "fields" LabelAxis (e.g. ['moon', 'sun']).
  • get_cuts(aman, sidelobes="sidelobes", object_list=None,
    rise_tol=1 deg) returns one RangesMatrix [ndet, nsamp] per object.
    Clear errors if the metadata is missing or the map has no WCS.
  • make_cuts: skips objects below rise_tol at the first, middle and
    last sample; projects the mask with
    P.for_tod(instrument_centered=...); uses unit responsivities so the
    mask value is independent of the detector T response (the old
    astype(int) could round 1*T < 1 down to 0); cuts samples where the
    projected mask is non-zero.
• site_pipeline/make_ml_map.py: call sidelobes.get_cuts(obs) with the
  new API. The --sun-mask/--moon-mask options are no longer used.
• tests/test_pmat.py: test_pmat_instrument_centered checks gamma
  independence, that the angular distance from the origin equals the
  detector-source distance on the sky, the orientation for the
  boresight detector against the Moon's az/el, and that
  from_map/to_map run; comps='TQU' is rejected.

Validation on a LAT observation (el = 120 deg, roll = 60 deg): the
spread across detectors at fixed time is ~0.6 x 0.5 deg (one wafer),
and the boresight orientation check agrees with skyfield's Moon az/el
to ~1e-4 at all tested samples.

WARNING

This implementation can ONLY be used to go from a mask to flagged
samples, "T only". It CANNOT be used to make maps in
instrument-centered coordinates: polarization is not supported
(gamma is set to 0 and only comps='T' is allowed), and mapmaking in
these coordinates has not been designed or validated. That will need a
future implementation.

Known open point: the direction in which detector offsets are rotated
by the roll (+roll vs -roll) has not been verified independently; for a
~1 deg LAT wafer the effect is at most ~1 deg.

Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com

[chervias]

this will fail tests without the so3g PR, so we need to merge that first