Fix CPU Huffman truncation of wide quantization outliers

[VarShankar]

Summary

The MGARD-CPU Huffman path accepts quantized coefficients as long int, but its outlier (miss) stream narrows every value to int. On LP64 systems, valid quantization indices outside the signed 32-bit range are therefore silently truncated, and decompression can violate the requested error tolerance by orders of magnitude.

This change:

• keeps legacy int-width outlier streams when all misses fit in int;
• stores the complete miss stream as long int when any miss requires the wider type;
• decodes both legacy and wide streams using the recorded miss-byte count; and
• rejects inconsistent miss-stream sizes.

Ordinary existing streams retain their previous representation. A new regression test covers both the legacy-width path and values outside the int range.

Validation

• The focused reproducer fails on unpatched master and passes with this change.
• The new Catch2 regression test passes.
• All 310 configured MGARD-CPU and serial MGARD-X tests pass.

[JieyangChen7]

@VarShankar Thanks a lot for your contribution to fixing this issue for MGARD-CPU. We are going to merge your PR.

[VarShankar]

Of course! I ran into this in a problem I was working on, so I figured I should do a PR. Thanks for maintaining MGARD!