Replace keystone-engine with a minimal in-house Thumb-2 assembler

[BrianPugh]

keystone-engine is unmaintained and painful to build (no macOS-arm64 wheel), and the patcher only ever assembles a small, fixed set of Thumb-2 instructions. Replace it with a purpose-built, dependency-free assembler.

• Add gnw_patch/thumb_asm.py: encodes movw (T3), mov.w/add.w/sub.w (T2/T3 modified immediate, incl. conditional forms), mov (T1), sub sp (T2), ldr.w [pc, #imm] (T2 literal), narrow/wide b (T2/T4), and IT blocks. Includes the ARM ThumbExpandImm encoder. Rejects operands keystone rejects.
• Wire it into FirmwarePatchMixin.asm(); drop CachedKeystone, the lazy keystone import, the keystone_cache.json hack, compact_json_encoder.py, and the dead InvalidAsmError.
• Move keystone-engine from the optional [patch] group to dev (Linux/Windows only via a platform marker) where it serves purely as the test oracle.
• Add tests/test_thumb_asm.py: compares assemble() against keystone-engine exhaustively (every movw immediate, every modified immediate, branch sweeps, IT blocks, register guards) plus the exact instruction strings the Mario/Zelda patches emit. Skips gracefully where keystone is unavailable.

Also covers the assembly forms introduced by PR #381 (ldr.w pc-literal with signed offsets and conditional add.w), verified byte-for-byte against keystone.