Improve SDL frame pacing with presentation-aware timing and conservative backend fallbacks

[Copilot]

SDLComponentNative was deriving animation/update deltas from CPU render-start time, which can diverge from the cadence at which frames become visible. This change adds presentation-aware timing and optional frame pacing, while keeping defaults conservative and falling back cleanly when a backend cannot provide reliable display-synchronized feedback.

• Public pacing configuration

  • Added ComponentNative::FramePacingMode with:
    • automatic
    • off
    • software
    • presentationDriven
  • Added orthogonal ComponentNative::Options controls:
    • withFramePacingMode(...)
    • withMaximumFramesInFlight(...)
  • Kept default behavior conservative: no forced low queue depth, and automatic mode only uses presentation-driven pacing when the backend can support it safely.

• Backend-neutral pacing policy

  • Introduced an internal FramePacer to centralize:
    • effective mode selection
    • software wake deadlines
    • adaptive lead-time estimation
    • presentation-vs-CPU delta selection
    • reset behavior across minimize/focus/resize/resume/surface-loss paths
  • The pacer distinguishes:
    • CPU update/render duration
    • submission duration
    • GPU completion telemetry
    • validated presentation timing
  • Missed deadlines advance to the next future slot instead of generating catch-up bursts.

• SDL window/render-loop integration

  • Refactored SDLComponentNative render scheduling to use FramePacer.
  • internalRefreshDisplay() now consumes presentation-derived deltas when validated timing exists, otherwise falls back to monotonic CPU time.
  • Added lifecycle resets so stale timing does not leak across:
    • minimize / restore
    • focus transitions
    • resize / display change
    • device-reset / surface-loss scenarios
  • Preserved repaint-on-demand semantics: no idle frames are generated just to harvest timing samples.

• GraphicsContext timing/capability hooks

  • Extended GraphicsContext with minimal backend-neutral timing support:
    • timing capability query
    • latest timing sample query
    • optional frame-latency wait
    • optional max-frames-in-flight control
  • Capability reporting is separate from sample validity, so callers can see whether a path is supported without conflating it with the latest usable timestamp.

• Backend support

  • Direct3D 11
    • Added DXGI frame-latency wait support where available.
    • Added DXGI frame statistics sampling for validated presentation timestamps.
    • Added max-frames-in-flight propagation with conservative clamping/reset behavior.
    • Cleans up disjoint/invalid samples instead of reusing stale presentation data.
  • Metal
    • Added max drawable count control when supported by platform availability.
    • Added asynchronous GPU completion telemetry.
    • Added presented-time sampling via drawable presentation callbacks where available, without treating GPU completion as display time.
  • OpenGL / SDL
    • Keeps conservative fallback behavior.
    • Uses software pacing / present-block telemetry only; does not treat swap return or GL fences as presentation timestamps.

• Bindings and tests

  • Exposed the new pacing options in Python bindings.
  • Exposed timing capability/info structs and accessors in graphics bindings.
  • Added deterministic pacing tests covering:
    • automatic/fallback mode selection
    • presentation-derived stable deltas
    • adaptive lead time
    • no catch-up burst after overruns
    • non-integral target cadence behavior
    • invalid/missing feedback fallback
    • reset semantics
    • stale/out-of-order presentation sample handling

Example:

auto options = ComponentNative::Options()
    .withVSync (true)
    .withFramePacingMode (ComponentNative::FramePacingMode::presentationDriven)
    .withMaximumFramesInFlight (2);