Move cold opcode bodies out of the VM dispatch loop

[hadashiA]

Summary

Shrinks the interpreter's giant dispatch method by moving cold opcode bodies into [MethodImpl(NoInlining)] local functions, following the extraction pattern already used for Class/Module/Exec/Rescue/RaiseIf etc.

Extracted

• OP_ARGARY — whole body
• OP_KARG — whole body
• OP_HASH / OP_HASHADD — literal-building loops
• SendInternal's method-visibility validation (the check stays inline; raise-path arg building moves out)
• EnterSlowPath — was an un-annotated local function; now pinned NoInlining so the JIT can never inline its ~150 lines back into the dispatch loop

Execute IL: 13,220 → 12,235 bytes; Tier1 native code: 26.5KB → 24.1KB (−9%).

Numbers (order-alternated pairs, within-pair comparison, vs #191)

benchmark	before	after
optcarrot	23.9–24.0 fps	24.4–24.5 fps (+2.2%), checksum 59662
bm_ao_render	~2735 ms	~2597 ms (−5.1%)
bm_ivar	—	neutral
bm_so_mandelbrot	~815 ms	~888 ms (+9%) — known regression

The mandelbrot regression is a code-placement effect on the Float fast path of Add..Div, not an execution-path change: bisecting showed even extracting only OP_ARGARY (which never executes in mandelbrot) reproduces it, and four different extraction combinations all land at the same ~+9%. optcarrot and ao improve consistently across all rounds. Mandelbrot-style float loops are the workload where the mrb2cs AOT path already delivers ~16x, so the interpreter trade-off favors the real workloads.

Tests

157 passed; all TFMs build; optcarrot video checksum identical.

🤖 Generated with Claude Code