Higher-level engines like macroquad issue one begin_pass / end_render_pass pair per draw call so they can stream batched draws to whatever the active camera's render target is. Each pair was becoming its own MTLRenderCommandEncoder, which on tile-based GPUs forces a full color-attachment store + (for MSAA) resolve on every draw — heavy memory bandwidth that visibly tanks real-device frame rate at retina resolution with Conf.sample_count > 1.

OpenGL absorbs this pattern transparently because re-binding an FBO doesn't trigger a resolve; Metal's per-encoder load/store actions make the cost explicit, and minimizing encoder count is a standard Metal best practice for exactly this reason.

This PR defers endEncoding until the next begin_pass actually requires a new encoder (different target, Clear action, or commit_frame). If the next begin_pass continues the same target with PassAction::Nothing, it keeps encoding into the existing encoder so the whole sequence collapses into one Metal pass with a single store/resolve at the real end. Each existing draw call's apply_pipeline / apply_bindings / apply_uniforms already re-sends its full state, so reusing the encoder is correctness-safe.

Tested on iPhone running iOS 27 with macroquad's per-draw-call begin/end pattern at 4× MSAA — drops the per-frame encoder count from the draw-call count down to one per render target, eliminating real-device frame lag.