Fix undefined behavior in Huffman encoder for single-symbol input

[alexey-milovidov]

Problem

HuffmanEncoder::build_code is invoked with len == 0 for the tree root (see the build_code(huffmanTree->qq[1], 0, 0, 0) call). When the Huffman tree has a single symbol — which happens for constant input data, where every value maps to the same quantization code — the leaf is the root, so the code length is zero and the encoder executes:

(huffmanTree->code[n->c])[0] = out1 << (64 - len);  // out1 << 64 when len == 0

Shifting a 64-bit value by 64 bits is undefined behavior in C++. It aborts under UndefinedBehaviorSanitizer (shift exponent 64 is too large for 64-bit type) and is generally unsafe.

Fix

Handle len == 0 explicitly (a zero-length code is empty, store 0). Also add a symmetric guard for len >= 128 in the other branch, where out2 << (128 - len) would shift by >= 64 (such codes do not fit in 128 bits anyway).

Context

This was found while integrating SZ3 into ClickHouse, where compressing a column of constant floating-point values reliably triggered the crash on a UBSan build: ClickHouse/ClickHouse#108788