perf: Use same reallocation strategy as Vec

This change seems to improve performance on some benchmarks (e.g.
techempower), but causes performance regression on another benchmark. We
need to measure performance using more test cases.

sailfish/src/runtime/buffer.rs

@@ -162,11 +162,10 @@ impl Buffer {
     }
 
     #[cfg_attr(feature = "perf-inline", inline)]
-    #[cold]
     fn reserve_internal(&mut self, size: usize) {
         debug_assert!(size <= std::isize::MAX as usize);
 
-        let new_capacity = std::cmp::max(self.capacity * 2, self.capacity + size);
+        let new_capacity = std::cmp::max(self.capacity * 2, self.len + size);
         debug_assert!(new_capacity > self.capacity);
         self.data = unsafe { safe_realloc(self.data, self.capacity, new_capacity) };
         self.capacity = new_capacity;