Add support for UPAL: joint point-line detection and description

[B1ueber2y]

Integrates UPAL (ECCV 2026) as a line
detector, a line descriptor/matcher, and — via the previously unimplemented
use_joint_point_line_detection path — a frontend that produces keypoints,
keypoint descriptors, line segments and line descriptors from a single
network pass per image.

Benchmarks

Detection and description only (no matching), 800x600 Hypersim images, RTX 3080:

stage	ms/image	output
deeplsd detect + wireframe describe	165.1	244 lines
aliked-n16 point detection	16.2	650 kpts
default frontend, detection total	181.3
upal detect + describe (lines only)	33.4	160 lines
upal joint, one pass (points + lines)	35.7	160 lines + 4096 kpts

4.9x on lines, 5.1x for the whole detection stage — the joint pass replaces
ALIKED as well. Adding the point half to UPAL costs only 2.3 ms, because it is
one extra descriptor sampling on a feature map that already exists.

Quickstart reconstruction (Hypersim ai_001_001, 98 images, groups off):

run	point frontend	line frontend	points	track	reproj	lines
A	aliked (554 kpt) + LightGlue	deeplsd + wireframe + gluestick	4849	7.98	1.108	584
B	aliked (554 kpt) + LightGlue	upal	4828	8.01	1.110	410
C	upal (4096 kpt) + mutual NN	upal	28210	7.42	1.287	452
D	aliked (2289 kpt) + LightGlue	deeplsd + wireframe + gluestick	17982	8.38	1.230	630

A vs B is the controlled comparison: identical point pipeline, so track
length (7.98 -> 8.01) and reprojection error (1.108 -> 1.110) are unchanged.
The line frontend swap costs 30% fewer 3D lines for 4.9x less time.

C and D are density-matched against each other (runs A and C differ in keypoint
count, so their point totals are not comparable — ALIKED is threshold-based and
returns ~554 by default, UPAL is top-k and returns exactly its budget). At
comparable density UPAL trails on track length (7.42 vs 8.38) and lines
(452 vs 630), most plausibly because mutual NN is matched against LightGlue and
GlueStick — UPAL has no trained matcher.

What's added

• src/limap/image/joint_point_line/UPAL/ — detector and matcher. Both segments
  and endpoint descriptors come from one encoder pass; going through upstream's
  API would encode twice, since UPAL.describe_keypoints re-runs the backbone.
  A test asserts the single-pass descriptors match upstream's to 1e-5.
• src/limap/image/joint_point_line/ — JointPointLineDetectionOptions and the
  stage function, which writes the three artifacts the separate point and line
  paths produce (hloc .h5, segments, line descinfo), leaving everything
  downstream unchanged.
• BaseDetector.detect_and_extract_joint — the capability, declared alongside
  detect_and_extract. Line-only detectors are rejected up front.
• Registered as upal in get_detector / get_extractor / get_matcher, with
  a _UNCERTAINTY2D_DEFAULTS entry.
• cfgs/structure_triangulation/upal.yaml, README and docs entries.

Also fixes two pre-existing bugs that made the dual-function detector path
unreachable (detect_and_extract_all_images), committed separately.

Design notes

• Association derives its line matcher from joint_point_line_detection when
  joint detection is enabled, rather than requiring line_detection to repeat
  the method. Otherwise a missing extractor_method would silently build a
  wireframe extractor against UPAL descinfo.
• max_num_keypoints defaults to 4096. Note this is a hard count, not a cap:
  UPAL's detector is top-k with no score threshold, so every image returns
  exactly this many keypoints. It directly sets point density.
• UPALOptions lives in line/register_detector.py rather than beside UPAL,
  because DetectorOptions carries the field and moving it would create a
  module-level import cycle.

Dependencies

Installed separately, like HAWP / TP-LSD / LBD / RoMa / Progressive-X — see
misc/install/upal.md. Deliberately kept out of
requirements.txt: UPAL is GPU-only in practice.

• upal-local-features@git+... — the network. Not on PyPI.

Both are required: the network predicts keypoints, descriptors and a line
distance field but no segments; segments come from LSD seeded by those keypoints
and filtered by that field.

Weights are not shipped in the package; the checkpoint downloads on first use
from a pinned commit via the existing download_weights helper.
JointPointLineDetectionOptions.weight_path defaults to None, so the root
comes from weights_root() like every other detector (#161) rather than being
pinned here -- a test guards that, since pinning it would download the same
checkpoint into a second directory.

[B1ueber2y]

cc: @francois141 @RaphaelKreft