同一段代码，Int 在不同设备上位宽不同

for (name, device) in [("flex", Device::flex()), ("ndarray", Device::ndarray()), ("metal", metal.clone())] {
  let i = Tensor::<1, Int>::arange(0..3, &device);
  let s = Tensor::<1, Int>::full([4], 1_000_000_000, &device).sum();
  let data = s.into_data();
  println!(
    "{name:<8} arange {:?} | sum of 4 x 1e9: {:?} {:?} | to_vec::<i64>: {:?}",
    i.dtype(),
    data.dtype(),
    data.clone().convert::<i64>().to_vec::<i64>().unwrap(),
    data.to_vec::<i64>()
  );
}

flex     arange I32 | sum of 4 x 1e9: I32 [-294967296] | to_vec::<i64>: Err(DTypeMismatch { expected: I64, actual: I32 })
ndarray  arange I64 | sum of 4 x 1e9: I64 [4000000000] | to_vec::<i64>: Ok([4000000000])
metal    arange I32 | sum of 4 x 1e9: I32 [-294967296] | to_vec::<i64>: Err(DTypeMismatch { expected: I64, actual: I32 })
case int-dtype: done
exit=0

### int-big flex

thread 'main' (3524870) panicked at ~/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-std-0.22.1/src/element/cast.rs:218:1:
Element cannot be represented in the target type: "i64"(3000000000) => "i32"
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
exit=101
### int-big ndarray
ndarray: I64 [3000000000]
case int-big: done
exit=0
### int-big metal

thread 'main' (3524904) panicked at ~/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/burn-std-0.22.1/src/element/cast.rs:218:1:
Element cannot be represented in the target type: "i64"(3000000000) => "i32"
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
exit=101

for (name, device) in [("flex", Device::flex()), ("metal", metal.clone())] {
  let a = Tensor::<1>::from_data(TensorData::from([0.1f64, 0.2]), &device);
  println!("{name:<6} from f64 data -> {:?}, supports F64 {}", a.dtype(), device.supports_dtype(DType::F64));
}
let a = Tensor::<1>::from_data(TensorData::from([0.1f64, 0.2]), (&metal, DType::F64));
println!("metal explicit F64 -> {:?}", a.dtype());
println!("{:?}", a.into_data());

flex   from f64 data -> F32, supports F64 true
metal  from f64 data -> F32, supports F64 false
metal explicit F64 -> F64
TensorData { bytes: Bytes { data: [154, 153, 153, "..."], len: 16 }, shape: Shape { dims: [2] }, dtype: F64 }
case f64: done
exit=0

let a = Tensor::<1>::from_floats([0.1, 0.2], &metal).cast(FloatDType::F64);
println!("cast -> {:?}", a.dtype());
println!("{:?}", a.into_data());

cast -> F64

thread 'main' (3523777) panicked at src/main.rs:131:26:
Error while reading data: use `try_into_data` instead to catch the error at runtime: An error happened during execution
Caused by:
  Several failures at once
Caused by:
  The bytes were never written (failure #2, still claiming 2 buffer(s))
Caused by:
  A launch error happened
Caused by:
  A compilation error happened during launch
Caused by:
  An error caused the compilation to fail
Caused by:
  the device refused the MSL module of kernel `cast_element_i_f32_o_f64_n_2_0e1143f3`: Validation Error

Caused by:
  In Device::create_shader_module_passthrough, label = 'cast_element_i_f32_o_f64_n_2_0e1143f3'
    Failed to generate the backend-specific code
...
exit=101

let big = Tensor::<1, Int>::from_data([3_000_000_000i64], (&device, DType::I64));

flex     I64 [3000000000]
ndarray  I64 [3000000000]
metal    I64 [3000000000]
case int-explicit: done
exit=0

let mut device = Device::metal(DeviceKind::DefaultDevice);
println!("configure F64: {:?}", device.configure(FloatDType::F64));
println!("configure I64: {:?}", device.configure(burn::tensor::IntDType::I64));
let s = Tensor::<1, Int>::full([4], 1_000_000_000, &device).sum();
println!("sum {:?} {:?}", s.dtype(), s.into_data().to_vec::<i64>().unwrap());

configure F64: Err(UnsupportedDType { device: "Cube(Wgpu(WgpuDevice { kind: DefaultDevice, backend: Metal }))", dtype: F64 })
configure I64: Ok(())
sum I64 [4000000000]
case configure-int: done
exit=0

first tensor F32
configure after use: Err(AlreadyInitialized { device: "Flex(Cpu)" })
configure before use: Ok(())
then tensor F64
case configure: done
exit=0