Improve hsd_803B51C8

[g-guthrie]

What changed

• replace four fixed-size JPEG metadata memcpy calls with typed aggregate copies so MWCC emits the target's inline word-copy sequences
• hoist the quantization divisor once per coefficient loop instead of reloading the global for every coefficient
• limit the chroma quantization table pointer lifetime so it reuses the target register
• align the stack frame and width/height register allocation with the original function

Why

hsd_803B51C8 was one of the largest remaining fuzzy mismatches. The fixed-size copies and repeated global loads produced substantially different instruction sequences from the original JPEG encoder.

The local target-object comparison improves from 81.44211% to 95.6%.

Validation

• uvx pre-commit run --files src/sysdolphin/baselib/hsd_3B34.c
• exact MWCC 1.2.5n compile with project flags
• local objdiff comparison against the exported hsd_803B51C8 target object: 95.6%

[decomp-dev]

Report for GALE01 (7eb1511 - 403a22f)

📈 2 improvements in unmatched items

Unit	Item	Bytes	Before	After
main/sysdolphin/baselib/hsd_3B34	hsd_803B51C8	+425	80.44%	96.40%
main/sysdolphin/baselib/hsd_3B34	extab	+1	88.64%	90.91%