Clarify frame-model host timing contracts

[mstan]

Follow-up for #22.

Summary:

• clarify that interp_bridge_run_loop() and related bridge entries are not CPU-only helpers: interpreted opcodes advance the shared Snes beam
• clarify the current APU policy: ordinary SNES interpreter fallback still uses relative catch-up; only active SA-1 absolute timeline suppresses duplicate catch-up
• point frame-model hosts from the relevant bridge/deadline APIs to docs/FRAME_MODEL_HOSTS.md
• document snes_catchupApu() as a frequent-slice catch-up API rather than a frame-at-a-time pacing API
• fix the per-directory interp816 test runners so bridge_test is built with the same Tier2 test define/source as the top-level C runner

Validation:

• git diff --check
• tests/interp816/run.ps1
• tests/runtime_dispatch/run_apu_port_guest_time_test.ps1
• bash tests/run_c_tests.sh (passes; existing renderer warnings only)