Fix StereoFromMotion words3D indexing for RGB-only input

[Miller2018]

Fix StereoFromMotion 3D-word indexing on RGB-only input

Fixes #1797.

Problem

With Mem/StereoFromMotion=true on RGB-only input, Memory::createSignature filled words3D
via push_back while iterating the words multimap (word-id order), so words3D[i] did not
belong to kptIdx i. Every consumer (DBDriverSqlite3.cpp:4624, RegistrationVis) indexes
words3D by kptIdx, so loop-closure PnP always produced 0 inliers
(Not enough inliers 0/20 (matches=N)), making loop closure unusable.

Fix

In the words3DWasEmpty branch, pre-size words3D to words.size() with NaN, then use the
same words3D[iter->second] = jter->second write as the non-empty branch. No behavior change
for RGB-D/stereo input (non-empty path untouched).

if(words3DWasEmpty)
{
    words3D.resize(words.size(), cv::Point3f(bad_point,bad_point,bad_point));
}

Validation

• Real 12-min indoor RGB-only dataset (external VIO odometry, rtabmap 0.23.7):
  loop 3→63 PnP goes 0/97 → 48/65 inliers (0.65 px reprojection); the run logged
  Not enough inliers 0/20 (matches=43) before the fix.
• Root-cause probe + raw measurements can be shared on request (see issue).
• Patch applies cleanly on master and 0.23.7 (git apply --check).

[matlabbe]

LGTM! Thanks for the PR

The merge-base changed after approval.