Specialize generated byte storage accessors

[RejectKid]

What changed

This combines the remaining generated-accessor optimizations into one change:

• generate direct endian-aware access for byte-aligned Span<byte>, ReadOnlySpan<byte>, byte arrays, and byte InlineArrays
• generate direct Span/InlineArray boolean masks and branch-based writes
• eliminate MemoryMarshal.AsBytes for eligible byte InlineArray fields
• specialize 11-, 12-, 24-, and 48-bit integral/enum fields with guarded 32/64-bit windows
• retain no-inline exact-length fallbacks when a backing buffer is smaller than the preferred wide window
• expand permanent benchmark coverage and the original-performance regression set from 74 to 88 methods
• add randomized LSB/MSB, signed, read-only, InlineArray, and exact-length correctness coverage

Memory-backed booleans intentionally remain on BitPrimitives: focused measurements showed that repeatedly acquiring .Span made direct generated code slower. Span and byte-InlineArray booleans use the direct path.

Why

The generator knows storage kind, field width, offset, signedness, and bit order at compile time. Eligible layouts can avoid the general primitive dispatch/validation path while preserving the existing generated API and falling back safely for exact-size buffers.

Performance

Paired BenchmarkDotNet comparisons against merge commit 9179007:

• 14-method ShortRun geometric mean: 2.36% faster, no regression beyond 5%
• 7-method MediumRun decision set: 2.08% faster, no regression beyond 5%
• Span boolean setter: 9.03% faster (MediumRun, confirmed)
• 11-bit Memory getter: 5.20% faster (MediumRun)
• 12-bit Memory getter: 1.09% faster (MediumRun)
• 48-bit Memory getter: effectively unchanged (0.18% faster)
• all measured operations remain allocation-free

Correctness and compatibility

• no public API changes
• exact-length 2-byte and 7-byte buffers use bounded cold fallbacks
• both LSB and MSB paths are covered
• signed extraction and enum paths are covered
• read-only Span/Memory getters are covered

Validation

• dotnet test BitsKit.Tests/BitsKit.Tests.csproj -c Release --no-restore
  • 4,589 passed on .NET 8
  • 4,589 passed on .NET 10
• dotnet build BitsKit.sln -c Release --no-restore
• dotnet pack BitsKit/BitsKit.csproj -c Release --no-build --no-restore
• git diff --check
• paired ShortRun and MediumRun regression harnesses against 9179007