Improve rendering time by rewriting `STD140BufferFormat._write_padded…

…()` (#4058)

The original implementation of `STD140BufferFormat._write_padded()` used `np.pad` which is slow.
This new implementation avoids that by creating a new array of zeroes instead. In this way, this method now takes 60x less time, and the total rendering time is almost halved.

manim/renderer/buffers/buffer.py

@@ -50,7 +50,7 @@ def __init__(
         self.name = name
         self.binding = binding
         self.dtype = []
-        self._paddings = {}  # LUT for future writes
+        self._paddings: dict[str, int] = {}  # LUT for future writes
         byte_offset = 0  # Track the offset so we can calculate padding for alignment -- NOTE: use RenderDoc to debug
         for data_type, var_name in struct:
             _base_char, base_bytesize, shape = self._GL_DTYPES[data_type]
@@ -97,11 +97,16 @@ def _write_padded(self, data: tuple | np.ndarray, var: str) -> np.ndarray:
         np.ndarray
             the same data with 0 or 1 columns of 0s appended
         """
-        try:
-            # This fails for 1D data (python or np.array)
-            return np.pad(data, ((0, 0), (0, self._paddings[var])), mode="constant")
-        except Exception:
-            return np.pad(data, ((0, self._paddings[var])), mode="constant")
+        data = np.asarray(data)
+        if self._paddings[var] == 0 or data.ndim == 0:
+            return data
+
+        # Make a new array with extra columns of 0s
+        new_shape = list(data.shape)
+        new_shape[-1] += self._paddings[var]
+        padded_data = np.zeros(new_shape)
+        padded_data[..., : data.shape[-1]] = data
+        return padded_data
 
     def write(self, data: dict) -> None:
         """Write a dictionary of key value pairs to the STD140BufferFormat's data attribute