refactor(gpu): eliminar GPU compute boids (prevé crash macOS)

Elimina el kernel Metal O(N²) de boids en GPU que causava GPU timeout
a macOS amb >50K boles, arrossegant WindowServer fins al crash.

- Elimina gpu_boid_buffer.hpp/cpp (GpuBoidBuffer, BallComputeData, BoidParams)
- Elimina kBoidComputeMSL i kBallComputeVertMSL de gpu_pipeline
- Elimina boid_compute_pipeline_ i ball_compute_pipeline_
- Elimina use_gpu_boids_, boid_params_, ball_screen_uniforms_ de Engine
- Elimina syncAndExitGpuBoids() i tot el compute dispatch de render()
- Mode BOIDS ara usa sempre boid_manager_ (CPU, spatial hash O(N))
  i renderitza via gpu_ball_buffer_ instanced (mateix path que PHYSICS)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

source/gpu/gpu_ball_buffer.hpp

@@ -0,0 +1,47 @@
+#pragma once
+
+#include <SDL3/SDL_gpu.h>
+#include <cstdint>
+
+// ---------------------------------------------------------------------------
+// BallGPUData — 32-byte per-instance record stored in VRAM.
+// Positions and sizes pre-converted to NDC space on CPU so the vertex shader
+// needs no screen-dimension uniform.
+//   cx, cy : NDC center   (cx = (x + w/2)/sw*2-1,  cy = 1-(y+h/2)/sh*2)
+//   hw, hh : NDC half-size (hw = w/sw,  hh = h/sh, both positive)
+//   r,g,b,a: RGBA in [0,1]
+// ---------------------------------------------------------------------------
+struct BallGPUData {
+    float cx, cy;      // NDC center
+    float hw, hh;      // NDC half-size (positive)
+    float r, g, b, a;  // RGBA color [0,1]
+};
+static_assert(sizeof(BallGPUData) == 32, "BallGPUData must be 32 bytes");
+
+// ============================================================================
+// GpuBallBuffer — owns the GPU vertex buffer used for instanced ball rendering.
+//
+// Usage per frame:
+//   buffer.upload(device, cmd, data, count);  // inside a copy pass
+//   // Then in render pass: bind buffer, SDL_DrawGPUPrimitives(pass, 6, count, 0, 0)
+// ============================================================================
+class GpuBallBuffer {
+public:
+    static constexpr int MAX_BALLS = 500000;
+
+    bool init(SDL_GPUDevice* device);
+    void destroy(SDL_GPUDevice* device);
+
+    // Upload ball array to GPU via an internal copy pass.
+    // count is clamped to MAX_BALLS.  Returns false on error or empty input.
+    bool upload(SDL_GPUDevice* device, SDL_GPUCommandBuffer* cmd,
+                const BallGPUData* data, int count);
+
+    SDL_GPUBuffer* buffer() const { return gpu_buf_; }
+    int count() const { return count_; }
+
+private:
+    SDL_GPUBuffer*         gpu_buf_      = nullptr;
+    SDL_GPUTransferBuffer* transfer_buf_ = nullptr;
+    int                    count_        = 0;
+};