Let an embedder observe guest execution and GX timing

[dougchansan]

Groundwork for ExpansionPak/ModernGekko#37, where every diagnostics zone reads 0.00 ms because nothing below the diagnostics library can report into it.

Retargeted from main to moderngekko-vendor. main does not configure for a ModernGekko consumer — it fails at Externals/tinygltf with No SOURCES given to target, with or without these changes (verified against plain main at 78bdce82). moderngekko-vendor is also the branch upstream ModernGekko pins.

Change

Three headers of optional pointers an embedder installs, null unless installed:

StaticRecompObserver.h — guest PC, guest execution time, and the core's existing dispatch/fallback/exception tallies.
VideoZoneObserver.h — FIFO, vertex, texture and shader time, plus draw/vertex/texture/EFB/shader tallies.
MmioObserver.h — MMIO read and write tallies, its own header so Core/HW gains no dependency on the others.

Probes at the points that already funnel: the dispatch loop, RunFifo, RunVertices, TexDecoder_Decode, the four ShaderCache::Compile*Shader methods, CopyRenderTargetToTexture, and MMIO Read/Write. 210 insertions, 1 deletion.

Cost

Granularity is the design. The guest PC is published on a 4096-dispatch cadence, guest time is measured once per timing slice, and the FIFO scope wraps a batch rather than each opcode. Cumulative counters are republished per slice rather than incremented per event. Measured rates that justify this: dispatches ~2.5M/s, MMIO ~26.6k/s, draw calls ~13.1k/s.

guest_cpu_ns is cumulative because the core has no notion of a frame; the consumer differences it.

Verification

Clean clone of ModernGekko's diagnostics branch with vendor/dolphin on this branch, x86_64 Linux, gcc 13.3.0, RelWithDebInfo:

CFG=0  BUILD=0  CTEST=0
100% tests passed, 0 tests failed out of 46

Mario Kart Double Dash GM4E01, windowed Vulkan, trace level:

GuestCpu                  13.861 ms      static_recomp_dispatches   52349514
ShaderGeneration           0.791 ms      mmio_reads/writes    313671 / 221773
GxCommandProcessor         0.356 ms      draw_calls                   280654
VertexLoader               0.024 ms      vertices_loaded             1752014
TextureDecoder             0.016 ms      efb_copies                     1028

LIKELY BOTTLENECK: CPU / guest execution bound
  Guest execution consumed 55.7% of frame time

Previously the zone table read (no zone timings recorded), every counter was 0, and the verdict was Insufficient evidence at 0.0% instrumented. hotspots.csv went from header-only to 512 rows, topped by 0x800EB638, which matches the core's own top dispatch site. Report contents remain free of usernames, paths and addresses.

Overhead across basic/detailed/trace is indistinguishable; repeated runs land within run-to-run host variance.

Not covered

Pipelines: CreatePipeline has 30 call sites and 8 backend implementations, so counting it needs the virtual wrapped in AbstractGfx — including D3D11/D3D12, which cannot be compiled or verified here. A documented zero beats an unverifiable cross-backend rename.

TextureDecodeBytes needs a per-format bits-per-pixel lookup rather than a tally, and GuestInstructions would need new counting on the 2.5M/s dispatch path. Both left out rather than approximated.

Note

The six VideoCommon and MMIO files are byte-identical between this branch and the revision ModernGekko pins, so they carry over as tested. StaticRecompCore_Run.cpp differs and was re-expressed against this branch's loop — that version is what the verification above was built from.