Fix bug in HostManagedBuffers logic

[Bastacyclop]

This didn't work, hitting an assertion:

depFun((n: Nat) => fun((n`.`i32) ->: (n`.`i32))(in =>
  mapSeq(add(li32(2)))(in) |> store(x =>
  oclRun(LocalSize(2), GlobalSize(n/2))(mapGlobal(add(li32(1)))(x)))
))

Now, it generates:

#include "ocl/ocl.h"
struct foo_t {
  Kernel k0;
};

typedef struct foo_t foo_t;

void foo_init(Context ctx, foo_t* self){
  (*self).k0 = loadKernel(ctx, k0);
}

void foo_destroy(Context ctx, foo_t* self){
  destroyKernel(ctx, (*self).k0);
}

void foo_run(Context ctx, foo_t* self, Buffer moutput, int n1, Buffer me2){
  {
    Buffer mx101 = createBuffer(ctx, n1 * sizeof(int32_t), HOST_WRITE | DEVICE_READ);
    {
      int32_t* x101 = (int32_t*)hostBufferSync(ctx, mx101, n1 * sizeof(int32_t), HOST_WRITE);
      int32_t* e2 = (int32_t*)hostBufferSync(ctx, me2, n1 * sizeof(int32_t), HOST_READ);
      /* mapSeq */
      for (int i_110 = 0; i_110 < n1; i_110 = 1 + i_110) {
        x101[i_110] = ((int32_t)2) + e2[i_110];
      }

    }

    {
      DeviceBuffer b0 = deviceBufferSync(ctx, moutput, n1 * sizeof(int32_t), DEVICE_WRITE);
      DeviceBuffer b2 = deviceBufferSync(ctx, mx101, n1 * sizeof(int32_t), DEVICE_READ);
      const size_t global_size[3] = (const size_t[3]){n1 / 2, 1, 1};
      const size_t local_size[3] = (const size_t[3]){2, 1, 1};
      const KernelArg args[3] = (const KernelArg[3]){KARG(b0), KARG(n1), KARG(b2)};
      launchKernel(ctx, (*self).k0, global_size, local_size, 3, args);
    }

    destroyBuffer(ctx, mx101);
  }

}

void foo_init_run(Context ctx, Buffer moutput, int n1, Buffer me2){
  foo_t foo;
  foo_init(ctx, &foo);
  foo_run(ctx, &foo, moutput, n1, me2);
  foo_destroy(ctx, &foo);
}

However there seems to be another type of bug remaining. It looks like the "one_copy" buffer implementation is failing to produce the correct output.