Fix MGARD-X HIP Huffman lossless (codec correctness + code-length OOB)

[gqian-coder]

Four independent bugs in the MGARD-X HIP ParallelHuffman lossless path: three in the Huffman codec (canonical-codeword generation, header serialization, and codeword-length overflow handling) made GPU-compressed streams undecodable/oversized or abort, and one out-of-bounds device read in the parallel code-length generation faulted on the GPU. All verified on gfx90a (Frontier MI250X/MI210): HIP-compressed streams now round-trip and are byte-compatible with the SERIAL decoder.

HuffmanWorkspace.hpp: allocate status as non-managed device memory.
It was managed (hipMallocManaged); atomicMin on fine-grained memory is
unreliable on ROCm/gfx90a, so GenerateCW::Operation4's atomic min never
updated newCDPI. The Huffman length-group boundary was never found, all
symbols collapsed into one group, and the canonical codewords came out
with a constant offset -- codebook and decodebook mutually inconsistent,
so the stream was undecodable by any backend.

Huffman.hpp: default-initialize outlier_count = 0.
ComposedLosslessCompressor::Compress -> CompressPrimary never writes
outlier_count, but Serialize reads it to size compressed_data.resize().
The uninitialized value produced a garbage-sized allocation (hipMalloc
OOM, or an invalid Copy1D).

GetCodebook.hpp: throw instead of exit(1) when the longest codeword
exceeds the H-type budget (sizeof(H)*8 - 8 bits), so callers can catch
it and fall back to another lossless backend or a smaller dict_size.

GenerateCL.hpp: bounds-check histogram[_lNodesCur + _curLeavesNum] in the
parallel code-length generation; when every remaining leaf joins the
batch the index reached dict_size (one past end of the nz_dict_size
histogram), faulting on GPU.