[H12Y10]

This PR supersedes #430 (which covers only the np.float fix) and completes a full
modern-Python compatibility chain for yolox/tracker, following up on #466 where the
three blockers are consolidated. It covers (1) the numpy>=1.24 alias removal, (2) the
C++ toolchain requirement, and (3) the lap wheel gap.

Changes

1. np.float removed in numpy >= 1.24 (matching.py, byte_tracker.py)
   Replace the removed alias so the tracker no longer raises
   AttributeError: module 'numpy' has no attribute 'float' on the first update().
   (Supersedes Removed deprecated np.float #430.)

2. C++ toolchain required by cython_bbox (cython_bbox_numpy.py, matching.py)
   matching.py uses exactly one function from cython_bbox (bbox_overlaps). This adds a
   pure-numpy implementation with the same semantics (including the original +1 pixel
   width/height convention) and imports it as a fallback through try/except ImportError,
   so inference-only use of yolox/tracker no longer requires MSVC / CppExtension - a
   common failure on Windows and on slim CI images. Behaviour is unchanged whenever
   cython_bbox is installed.

3. lap has no wheel on some platforms (README)
   Documented that on platforms where lap has no prebuilt wheel, the lapx distribution
   ships a module named lap (same lapjv API), so pip install lapx satisfies the
   existing import with no code change. No code path is modified for this item.

Verification

• The numpy alias restore and the pure-numpy bbox_overlaps shim are used in production in
  a downstream monocular AEB/FCW pipeline, covered by 41 zero-data regression tests (pytest collect) and CI
  on Ubuntu/Windows x Python 3.10/3.12: https://github.com/H12Y10/mono-aeb
• The numpy implementation was checked against the shim used there on randomized box sets -
  identical IoU matrices, including the +1 pixel convention and zero-overlap cases.
• For the lap note: pip install lapx was verified to install a module named lap
  (import lapx fails as expected), and lapjv assignment results were identical to lap
  on the cost matrices we validated (same total cost, same x/y assignment).

License

Same MIT as the repository; the changes are minimal and do not alter tracking behaviour.