Reduce STU parsing and schema initialization overhead

[Js41637]

STU loading repeatedly performs reflection, boxes primitive array elements, and rebuilds inherited field metadata. This change reduces that work through three optimizations:

• Bulk primitive-array reads: Read supported numeric arrays directly into typed storage instead of decoding and assigning each element individually. Reads use bounded chunks and handle partial stream reads. Custom readers and factories retain the existing decoding path.
• Cached struct readers: Cache a reader delegate for each struct type, replacing repeated generic method construction and reflective invocation for scalar fields and array elements. The cache is thread-safe and preserves the existing exception wrapping.
• Cached schema metadata: Enumerate assembly types once during initialization and reuse declared field attributes and inherited field ordering. Published attributes and arrays remain independently mutable, and custom attribute subclasses retain their reflection behavior.

Performance

Measured results for the implemented optimizations:

Measurement	Before	After	Reduction
STU parsing time	2.689 s	1.950 s	27.5%
Asset opening and parsing	4.139 s	3.365 s	18.7%
Parser allocations	3.792 GB	2.376 GB	37.3%
Cold schema initialization	180.05 ms	108.90 ms	39.5%
Schema initialization allocations	70.84 MB	40.80 MB	42.4%

The parsing results measure primitive-array and struct-reader improvements together across 186,596 assets, using the median of six measured passes after warmup. Schema initialization was measured separately across three fresh-process runs. Allocation figures represent cumulative allocated bytes, not peak memory usage.