Fix KNN query capacity for decoder interpolation

[mulyoved]

Summary

• raise the KNN query kernel capacity from 100 to the decoder's required 256 neighbors
• reject unsupported neighbor counts before CUDA launch
• add GPU regression coverage for 256-neighbor success and 257-neighbor rejection

Root cause

The decoder requests 256 neighbors while the CUDA kernel allocates fixed arrays for 100. The kernel writes past both arrays, corrupts returned indexes, and later triggers a PyTorch index-out-of-bounds assertion in interpolation. Clamping the corrupted indexes would hide the memory error, so this change fixes the capacity mismatch at its source.

Verification

• the 256-neighbor test failed on the unmodified upstream commit for the reproduced invalid-index reason
• both focused GPU tests pass after the fix on CUDA 11.1 and PyTorch 1.10
• the released Phase 1 sample completes inference, produces the expected NumPy output contract, and renders successfully on a Tesla T4

Fixes #7