Refactor the metal backend to always reuse command encoders/buffers unless a shared memory access is requested

candle-core/benches/benchmarks/mod.rs

@@ -25,7 +25,7 @@ impl BenchDevice for Device {
             }
             Device::Metal(device) => {
                 #[cfg(feature = "metal")]
-                return Ok(device.wait_until_completed()?);
+                return Ok(device.synchronize()?);
                 #[cfg(not(feature = "metal"))]
                 panic!("Metal device without metal feature enabled: {:?}", device)
             }

candle-core/src/convert.rs

@@ -1,7 +1,6 @@
 //! Implement conversion traits for tensors
 use crate::{DType, Device, Error, Tensor, WithDType};
 use half::{bf16, f16, slice::HalfFloatSliceExt};
-use std::convert::TryFrom;
 
 impl<T: WithDType> TryFrom<&Tensor> for Vec<T> {
     type Error = Error;