Set up cava benchmark

• Set up cava benchmark w/ criterion.
• GPU schedule for cava.
• Convert a bunch of benchmarks from using hercules refs to use hercules boxes.
• Define AlignedAlloc global allocator in hercules_rt that aligns everything to 32 bytes.
• Bump nom version.

hercules_rt/src/lib.rs

-#![feature(once_cell_try)]
+#![feature(once_cell_try, pointer_is_aligned_to)]
 
 use std::alloc::{alloc, dealloc, Layout};
 use std::marker::PhantomData;
@@ -189,6 +189,7 @@ pub struct CUDABox {
 
 impl<'a> HerculesCPURef<'a> {
     pub fn from_slice<T>(slice: &'a [T]) -> Self {
+        assert!(slice.as_ptr().is_aligned_to(32));
         let ptr = unsafe { NonNull::new_unchecked(slice.as_ptr() as *mut u8) };
         let size = slice.len() * size_of::<T>();
         Self {
@@ -201,7 +202,6 @@ impl<'a> HerculesCPURef<'a> {
     pub fn as_slice<T>(self) -> &'a [T] {
         let ptr = self.ptr.as_ptr() as *const T;
         assert_eq!(self.size % size_of::<T>(), 0);
-        assert!(ptr.is_aligned());
         unsafe { from_raw_parts(ptr, self.size / size_of::<T>()) }
     }
 
@@ -214,6 +214,7 @@ impl<'a> HerculesCPURef<'a> {
     }
 
     pub unsafe fn __from_parts(ptr: *mut u8, size: usize) -> Self {
+        assert!(ptr.is_aligned_to(32));
         Self {
             ptr: NonNull::new(ptr).unwrap(),
             size,
@@ -224,6 +225,7 @@ impl<'a> HerculesCPURef<'a> {
 
 impl<'a> HerculesCPURefMut<'a> {
     pub fn from_slice<T>(slice: &'a mut [T]) -> Self {
+        assert!(slice.as_ptr().is_aligned_to(32));
         let ptr = unsafe { NonNull::new_unchecked(slice.as_ptr() as *mut u8) };
         let size = slice.len() * size_of::<T>();
         Self {
@@ -257,6 +259,7 @@ impl<'a> HerculesCPURefMut<'a> {
     }
 
     pub unsafe fn __from_parts(ptr: *mut u8, size: usize) -> Self {
+        assert!(ptr.is_aligned_to(32));
         Self {
             ptr: NonNull::new(ptr).unwrap(),
             size,
@@ -268,6 +271,7 @@ impl<'a> HerculesCPURefMut<'a> {
 #[cfg(feature = "cuda")]
 impl<'a> HerculesCUDARef<'a> {
     pub fn to_cpu_ref<'b, T>(self, dst: &'b mut [T]) -> HerculesCPURefMut<'b> {
+        assert!(dst.as_ptr().is_aligned_to(32));
         unsafe {
             let size = self.size;
             assert_eq!(size, dst.len() * size_of::<T>());
@@ -309,6 +313,7 @@ impl<'a> HerculesCUDARefMut<'a> {
     }
 
     pub fn to_cpu_ref<'b, T>(self, dst: &mut [T]) -> HerculesCPURefMut<'b> {
+        assert!(dst.as_ptr().is_aligned_to(32));
         unsafe {
             let size = self.size;
             let ptr = NonNull::new(dst.as_ptr() as *mut u8).unwrap();