Fix remaining_length accounting after Huffman decode in regression/composed predictors

[groeneai]

Fixes a remaining_length byte-accounting bug that causes silent data loss for the ALGO_LORENZO_REG compression algorithm when used via the ClickHouse SZ3 codec (ClickHouse/ClickHouse#110336).

Root cause

HuffmanEncoder::decode() advances the read pointer past the encoded stream but does not update the caller-owned remaining_length bound. Callers must account for the consumed bytes themselves.

RegressionPredictor::load() did this incorrectly:

regression_coeff_quant_inds = encoder.decode(c, coeff_size);
encoder.postprocess_decode();
remaining_length -= coeff_size * sizeof(int);   // WRONG

coeff_size * sizeof(int) is the uncompressed index count in bytes, but decode() only advanced c by the Huffman-compressed stream size (sizeof(size_t) + encodedLength). Since the compressed stream is smaller, this over-subtracts and understates remaining_length for the following main-quantizer encoder.load(), whose tree bound check tree_bytes > remaining_length then spuriously throws SZ3 Huffman: tree exceeds compressed buffer.

RegressionPredictor is only used by ALGO_LORENZO_REG, so ALGO_INTERP/ALGO_INTERP_LORENZO are unaffected. Small element counts fail deterministically; large counts happen to leave enough slack in the main stream to survive the understated bound (in ClickHouse this shows up as: n=500 fails, n=5000 works).

Fix

Account for exactly the bytes decode() consumed, via pointer difference:

const uchar *decode_start = c;
regression_coeff_quant_inds = encoder.decode(c, coeff_size);
encoder.postprocess_decode();
remaining_length -= static_cast<size_t>(c - decode_start);

ComposedPredictor::load() had the mirror defect (it decoded its selection stream but never decremented remaining_length at all, leaving a loose bound). Tightened the same way.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
2 out of 5 committers have signed the CLA.

✅ alexey-milovidov
✅ groeneai
❌ ayzk
❌ gsylvand
❌ guoxiliu
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[groeneai]

cc @alexey-milovidov @scanhex12 — could you review? This fixes the ALGO_LORENZO_REG data-loss reported in ClickHouse/ClickHouse#110336 (small element counts fail every read with SZ3 Huffman: tree exceeds compressed buffer). The remaining_length bound was decremented by the uncompressed index count instead of the compressed stream size after Huffman decode. Once merged I'll bump the ClickHouse submodule and add a regression test.