Optimize YDR IO

[Hancapo]

Summary

• Materialize nominal vectors directly in the native decoder and reuse float64 position snapshots for bounds and vertex packing.
• Prepare drawable bounds once per save, including shared YDD/YFT writers; transform rigid geometry through shared NumPy math without temporary vectors.
• Remove redundant channel copies and index normalization, and resolve vertex declarations once per channel.
• Retain zlib (YDR level 6) and Python 3.11+ / abi3 compatibility.

Verification

• 2,114 unit tests and 7 subtests passed locally.
• 34 tests passed on an isolated installed wheel in Python 3.11, including the explicit ABI-floor test.
• 14 benchmark outputs are byte-identical to the previous optimization baseline in the same interpreter.
• Latest Python 3.14 writing improvements: 8,192-triangle grid 11.16 to 7.07 ms; multiple materials 22.59 to 13.69 ms; rigid skeleton 23.95 to 7.10 ms; skinning 14.66 to 10.77 ms.