Fix tile rendering skew from lidar outliers (#11)

[endolith]

Summary

Addresses issue #11: a few extreme Z values could skew DSM/high-pass range so slippy tiles rendered flat white.

Changes

1. Noise classes at ingest — When filterNoise is on, the PDAL pipeline now uses a streamable filters.expression stage to drop ASPRS Classification 7 (low noise) and 18 (high noise), not only class 18. Class 7 is the LAS “low point / noise” bucket that matches the reported very-low spikes; filters.expression stays compatible with execute_streaming.

2. Robust tile contrast — Before imshow, finite values are summarized with the 0.5 and 99.5 percentiles to set vmin/vmax, so a handful of remaining outliers cannot dominate matplotlib’s autoscale. Degenerate cases (empty or constant data) fall back safely.

Notes

• Cached .laz tiles built under the old pipeline still contain old classifications until re-fetched; new fetches get the stricter noise filter. The percentile scaling still improves PNGs from existing DSMs with rare bad pixels.
• Users who rely on class 7 points being present should turn off filterNoise (would require a code path exposing that flag to HTTP if desired later).