Fix: load_pc resize -> reserve

[bloom256]

Problem

PointCloud::load_pc calls resize(npoints) on points_local, intensities,
and timestamps, then appends real data with emplace_back. This creates
2*npoints elements: first half zero-initialized, second half real data.
std::vector::resize value-initializes new elements, which for numeric types
(double, unsigned short) means zero.

Why this bug is invisible

Phantom points have [0, 0, 0] local coordinates and timestamp == 0:

• With downsampling: all phantom points fall into a single bucket at origin
  and are reduced to one point. No practical effect.
• During export: points with timestamp == 0.0 are filtered out when
  skip_ts_0 is enabled (export_laz.h:418-421). Phantom points never reach
  the output file.

This bug affects only memory usage when point clouds are loaded without
downsampling.

Measurements

Dataset: 60.9M points, loaded without downsampling.

	Before (resize)	After (reserve)
Total points loaded	121,857,980	60,928,990
Peak during loading	11.3 GB	6.9 GB
Memory after loading	5.3 GB	3.4 GB

Fix

resize -> reserve on 3 lines in core/src/point_cloud.cpp.