Avoid excess copies in fletcher32

During encoding preallocate a `bytes` object for the final result and
write everything directly into it. This avoids unnecessary staging and
copying of intermediate results.

Make use of Cython typed-`memoryview`s throughout encode and decode for
efficient access of the underlying data.

Further leverage the `store_le32` and `load_le32` functions to quickly
pack and unpack little-endian 32-bit unsigned integers from buffers
when encoding and decoding.

Author: jakirkham

numcodecs/fletcher32.pyx

@@ -1,13 +1,14 @@
 # cython: language_level=3
 # cython: overflowcheck=False
 # cython: cdivision=True
-import struct
 
 from numcodecs.abc import Codec
 from numcodecs.compat import ensure_contiguous_ndarray
 
 from libc.stdint cimport uint8_t, uint16_t, uint32_t
 
+from ._utils cimport store_le32, load_le32
+
 
 cdef uint32_t _fletcher32(const uint8_t[::1] _data):
     # converted from
@@ -64,22 +65,34 @@ class Fletcher32(Codec):
         """Return buffer plus 4-byte fletcher checksum"""
         buf = ensure_contiguous_ndarray(buf).ravel().view('uint8')
         cdef const uint8_t[::1] b_mv = buf
-        val = _fletcher32(b_mv)
-        return buf.tobytes() + struct.pack("<I", val)
+        cdef Py_ssize_t b_len = len(b_mv)
+
+        cdef Py_ssize_t out_len = b_len + 4
+        cdef bytes out = PyBytes_FromStringAndSize(NULL, out_len)
+        cdef uint8_t* out_ptr = <uint8_t*>out
+        cdef uint8_t[::1] out_mv = (<uint8_t[:(out_len + 1):1]>out_ptr)[:out_len]
+
+        out_mv[:-4] = b_mv
+        store_le32(&out_mv[-4], _fletcher32(b_mv))
+
+        return out
 
     def decode(self, buf, out=None):
         """Check fletcher checksum, and return buffer without it"""
         b = ensure_contiguous_ndarray(buf).view('uint8')
-        cdef const uint8_t[::1] b_mv = b[:-4]
-        val = _fletcher32(b_mv)
-        found = b[-4:].view("<u4")[0]
+        cdef const uint8_t[::1] b_mv = b
+        val = _fletcher32(b_mv[:-4])
+        found = load_le32(&b_mv[-4])
         if val != found:
             raise RuntimeError(
                 f"The fletcher32 checksum of the data ({val}) did not"
                 f" match the expected checksum ({found}).\n"
                 "This could be a sign that the data has been corrupted."
             )
+
+        cdef uint8_t[::1] out_mv
         if out is not None:
-            out.view("uint8")[:] = b[:-4]
+            out_mv = ensure_contiguous_ndarray(out).view("uint8")
+            out_mv[:] = b_mv[:-4]
             return out
         return memoryview(b[:-4])