Optimize generated scalar getters

[RejectKid]

What changed

• emit direct mask-and-shift expressions for generated integral-backed LSB and MSB getters, including integral, enum, and Boolean properties
• fix Boolean getters backed by signed integral fields so a set one-bit field reads as true
• expand generated-accessor benchmarks to cover signed and unsigned integers, Boolean, enum, MSB, Memory<byte>, and inline-array models
• compare equivalent models compiled by the original and current source generators in the performance-regression harness, with focused method filtering
• update benchmark and contributor documentation

Why

Generated scalar getters previously delegated to general-purpose BitPrimitives methods even though their offsets, widths, orders, and backing types are compile-time constants. Emitting the operation directly gives the JIT a smaller accessor and avoids an extra helper call in Tier-0 code. It also fixes the signed one-bit Boolean case, where comparing a sign-extended -1 result with 1 returned false.

For the representative generated UInt32 MSB getter, disassembly changed from a 45-byte accessor that called BitPrimitives.ReadUInt32MSB to a 25-byte, call-free bswap/shift sequence. Tier-1 throughput remains effectively tied because the JIT already inlined the old helper.

Validation

• dotnet build BitsKit.sln -c Release --no-restore — succeeded with 0 warnings
• dotnet test BitsKit.Tests/BitsKit.Tests.csproj -c Release --framework net8.0 — 4,587 passed
• dotnet test BitsKit.Tests/BitsKit.Tests.csproj -c Release --framework net10.0 — 4,587 passed
• focused Short generated-accessor comparison — 10 operations, geometric mean 0.29% faster, no regression beyond 5%
• full Dry original/current harness validation — all 66 comparable operations discovered and executed