Improve lidar odometry performance by replacing mutexes with thread-local reduction

[bloom256]

This PR improves performance of lidar odometry compute_step_2 by removing mutex-protected shared reductions.

Changes:

• Replace mutex-protected shared accumulation with thread-local accumulation using tbb::combinable.
• Use tbb::blocked_range to amortize access to the shared combinable object, accessing it once per block instead of once per point.

Performance:

• ~25% speedup observed for compute_step_2 on an Intel Core i7 machine.

Notes:

• Intended to be behavior-preserving; only the parallel reduction mechanism changed.