Add det_left mode to projection engine for instrument-centered pointing

[chervias]

This change was implemented with Claude Code (Opus 5.5).

Motivation

The projection engine always combines boresight and detector quaternions as q_bore[t] * q_det[d]. Instrument-centered coordinates (the position of a source such as the Moon or Sun relative to each detector, on axes fixed to the telescope structure) need

q[d,t] = A[d] * X[t]
A[d]   = q_to_eq * Rz(roll0) * q_det[d]^-1 * Rz(-roll0)
X[t]   = Rz(roll[t]) * q_bore[t]^-1 * q_src[t]

where the per-detector factor sits on the LEFT of the per-sample one, so it cannot be expressed in the existing q_bore * q_det form. Rather than transposing the problem (which would give [nsamp, ndet] outputs and require transposed signal, cuts, threads and responses), this adds an opt-in flag that reverses the order of the product, so everything else (all projections, tiling, bilinear interpolation, HEALPix, threads, cuts, weights, [ndet, nsamp] layout) works unchanged.

C++ (src/Projection.cxx, include/Projection.h)

• Pointer gets a constructor Pointer(bool det_left=false) and a private generic helper GetQdet(i_time, dofs), which reads the boresight quaternion and returns q_ofs * q_bore if det_left is set, q_bore * q_ofs otherwise.
• The seven GetCoords specializations that combine quaternions (ProjQuat, ARC, SIN, TAN, ZEA, CEA, CAR) now call GetQdet instead of each duplicating the read-and-multiply block. The decoding of the combined quaternion into projection coordinates is unchanged. ProjFlat is untouched.
• ProjectionEngine gets a private _det_left member (default false) with get_det_left/set_det_left, and passes it to every Pointer it constructs (coords, pixels, pointing_matrix, pixel_ranges, tile_hits, tile_ranges, from_map, to_map, to_weight_map).
• EXPORT_ENGINE exposes it as the read/write Python property "det_left" on every ProjEng class, including the HEALPix engines. The default is False, so existing behaviour is unchanged.

Python (python/proj/)

• Assembly has a det_left attribute: Assembly.attach(sight, fplane, det_left=False).
• _ProjectionistBase.get_ProjEng(..., det_left=False) sets the property on the engine instance.
• New _ProjectionistBase._get_pointing(assembly) returns (q_bore, q_det, det_left) and is used by every projection method (get_pixels, get_pointing_matrix, get_coords, get_planar, to_map, to_weights, from_map, assign_threads, assign_threads_from_map, get_active_tiles, and ProjectionistHealpix.get_coords). The celestial-to-native rotation of the map must multiply the leftmost factor: normally it is applied to the boresight (cached q_native as before); with det_left it is applied to the detector quaternions.
• assign_threads(method='domdir') falls back to 'simple' for a det_left Assembly, since the domdir algorithm assumes the normal ordering.

Tests (test/test_proj_eng.py)

• test_40_det_left_engine: raw engine output with det_left off/on against numpy q_boreq_ofs and q_ofsq_bore.
• test_41_det_left_projectionist: a det_left Assembly gives the same pixels, coords, to_map and from_map results as an equivalent normal Assembly built per detector, on tiled and untiled TAN maps, and with the simple/domdir thread assignments.

WARNING

This is intended ONLY for projecting a temperature ("T only") mask into sample space in order to flag samples (e.g. Moon/Sun sidelobe cuts, see the companion sotodlib change). It must NOT be used to make maps in instrument-centered coordinates: the polarization angle is not meaningful in this mode (sotodlib builds the detector quaternions with gamma = 0 and only allows comps='T'), and mapmaking in these coordinates has not been designed or validated. That needs a future implementation.

[chervias]

Claude code session https://claude.ai/code/session_01TEzk2tan4q5HizPzkiqHqe