[r-bin]

--region-paced seems to have solved a lot of problems for me. This is an extension, so Evermore can use the native code as often as possible.

Summary

With --instruction-blocks --region-paced, compiled code checks the master-clock deadline before every instruction, so a CPU
slice can stop at any instruction. The scheduler then interprets the rest of the routine, and its callers after its return,
until the next call bounces. A host that slices often (DRAM refresh points, vblank edges, single steps around NMI enables)
ends up mostly interpreted: about 13 % of Secret of Evermore's instructions ran compiled.

--aot-resume gives each region-paced variant a resume switch, so code a slice stopped in goes on compiled:

 deadline reached in compiled code
   unwind to the scheduler
+  record each abandoned frame (S, entry S, return contract)
 scheduler at PC
-  interpret until the next call bounces
+  frame recorded at this S?  re-enter that variant at PC (cpu_dispatch_resume_paired)
+  no frame known?            enter as a frame one below S; its RTS/RTL hands back the popped PC

Off by default; the flag is part of the generation identity only when on (existing recipe digests do not change).
SNESRECOMP_AOT_RESUME=0 at run time turns it off, =2 keeps suspended frames only.

What keeps the tiers alike:

• An instruction compiled code hands back in place (unmodeled, MMIO, a pointer it cannot preview) runs once in the
  interpreter before resuming there again.
• The prologue's stack-window fallback hands back the resume point, not the variant's entry.
• A frameless frame's normal return (a bound routine that ran its own RTS/RTL) goes on at the address it popped.
• A compiled RTI goes on at the popped PC while the scheduler owns the run, so interrupt handlers resume too.
• The JSL envelope restores PB after a normal return even with an unwind pending.
• Each region-paced instruction that stays compiled reports itself to g_interp_bridge_pc_hook (cpu_aot_insn_begin),
  as the interpreter does.
• The beam is synced before compiled code reads PPU, joypad and $42xx registers.

Related: #276 adds yields at H/V IRQ latches to region-paced code; each such stop would otherwise hand the rest of the
routine to the interpreter, so the two complement each other. The compiled RTI here also covers what #251's Secret of Mana
run needed as a test-only addition (resumable interrupt handlers).

Evidence

Secret of Evermore (USA), two MesenCE 2.2.1 movies from power-on (8,592 and 23,756 NMIs), WRAM, VRAM, CGRAM and OAM
compared at every NMI, nothing copied or masked, on main 80a41241:

• Before (--instruction-blocks --region-paced): both replays identical; about 13 % of instructions compiled; the
  second replay runs in 93 s.
  After (--aot-resume added): both replays identical; 94.05 % / 92.43 % compiled; the second replay runs in 79 s
  (15 % faster end to end, PPU and APU included). Also identical with resume off at run time and all-interpreted.

pytest tests/v2 tests/test_snes_cycles.py: 638 passed. New: tests/v2/test_aot_resume.py (a resume entry per instruction,
off by default, needs region pacing, the fallback hands back the resume point, the resume table keeps the longest variant,
the flag check). The region-pacing harness gets stubs for the two resume helpers and cpu_aot_insn_begin.

Recipe: https://github.com/r-bin/SecretOfEvermoreRecomp/blob//docs/pr/README.md#reproduce

Merge Danger

Door: two-way. Opt-in at generation; off at run time with SNESRECOMP_AOT_RESUME=0.

Blast Radius: small for projects that do not pass --aot-resume. Two changes reach everyone using region pacing or the
bridge: region-paced instructions now call cpu_aot_insn_begin (a hook call, no timing change), and the beam is synced
before compiled register reads.

Who gains: projects that pass --region-paced and whose host stops the CPU often (per-line slices, single steps around
interrupts, IRQ-latch yields as in #276). With beam_frame_driver's few stops per field the gain is small; without region
pacing nothing changes.

Limits: measured on one game; the speed gain is one replay's wall time. Not yet run together with #276. Frameless entries
assume a frame one below S; a routine that returns from such an entry in an unusual way goes back to the interpreter at
the popped PC.