Improve hsd_803B51C8 matching

[g-guthrie]

What changed

• model the JPEG comment and Huffman tables as fixed-size aggregate types so MWCC emits the original inline copy sequences
• cache the quantization scale once per transformed block instead of reloading it for every coefficient
• align local layout, integer widths, and dimension declaration order with the target function

Why

hsd_803B51C8 used generic memcpy calls and repeatedly loaded the quantization scale inside three unrolled loops. Those source shapes differed substantially from the original compiler output.

The aggregate assignments reproduce the target's paired loads/stores and counted AC-table loops, while the cached scale reproduces the target's single load before each unrolled division block.

Matching improvement

• stripped function match: 81.44211% -> 94.7579%
• candidate size: 2592 -> 2664 bytes
• target size: 2660 bytes

Validation

• pre-commit run --files src/sysdolphin/baselib/hsd_3B34.c
• python3 tools/check/main.py --quiet src/sysdolphin/baselib/hsd_3B34.c
• git diff --check
• exact MWCC 1.2.5n compile and address-verified objdiff comparison against hsd_803B51C8

[decomp-dev]

Report for GALE01 (76a4770 - 664ec55)

📈 2 improvements in unmatched items

Unit	Item	Bytes	Before	After
main/sysdolphin/baselib/hsd_3B34	hsd_803B51C8	+402	80.44%	95.56%
main/sysdolphin/baselib/hsd_3B34	extab	+1	88.64%	90.91%

[g-guthrie]

Superseded by #3077, which improves the same function further (+425 bytes vs +402 bytes in the decomp.dev report). Keeping the stronger single contribution open.