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)]
 
-use std::alloc::{alloc, dealloc, Layout};
+use std::alloc::{alloc, dealloc, GlobalAlloc, Layout, System};
 use std::marker::PhantomData;
 use std::ptr::{copy_nonoverlapping, write_bytes, NonNull};
 use std::slice::{from_raw_parts, from_raw_parts_mut};
@@ -867,3 +867,24 @@ impl<'a, T> HerculesRefInto<'a> for Box<[T]> {
         HerculesCPURef::from_slice(self)
     }
 }
+
+/*
+ * We need all allocations to be aligned to LARGEST_ALIGNMENT bytes for
+ * vectorization. This is the easiest way to do that.
+ */
+pub struct AlignedAlloc;
+
+unsafe impl GlobalAlloc for AlignedAlloc {
+    unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
+        let layout = layout.align_to(LARGEST_ALIGNMENT).unwrap();
+        System.alloc(layout)
+    }
+
+    unsafe fn dealloc(&self, ptr: *mut u8, layout: Layout) {
+        let layout = layout.align_to(LARGEST_ALIGNMENT).unwrap();
+        System.dealloc(ptr, layout)
+    }
+}
+
+#[global_allocator]
+static A: AlignedAlloc = AlignedAlloc;