Address memory leak in imwrite_to_memory()

[claydugo]

imwrite_to_memory() leaks memory proportional to image size. Each call retains ~15 MB for a 3072×3072 image, leading to multi-gigabyte memory usage when encoding many images.

Minimal Reproducer

import gc
import numpy as np
import psutil
import ojph

def get_mem_mb():
    return psutil.Process().memory_info().rss / (1024**2)

initial = get_mem_mb()
buffers = []

for i in range(50):
    data = np.random.randint(0, 255, (2048, 2048), dtype='uint8')
    buffers.append(ojph.imwrite_to_memory(data, channel_order='HW'))

    if (i + 1) % 10 == 0:
        gc.collect()
        compressed_mb = sum(len(b) for b in buffers) / (1024**2)
        leaked = get_mem_mb() - initial - compressed_mb
        print(f"After {i+1} images: leaked={leaked:.0f} MB")

gc.collect()
compressed_mb = sum(len(b) for b in buffers) / (1024**2)
leaked = get_mem_mb() - initial - compressed_mb
print(f"\nTotal: compressed={compressed_mb:.0f} MB, leaked={leaked:.0f} MB")
print(f"Leak per image: {leaked/50:.1f} MB")

Expected output (before fix):

After 10 images: leaked=68 MB
After 20 images: leaked=154 MB
After 30 images: leaked=240 MB
After 40 images: leaked=325 MB
After 50 images: leaked=411 MB

Total: compressed=227 MB, leaked=411 MB
Leak per image: 8.2 MB

Expected output (after fix):

After 10 images: leaked=12 MB
After 20 images: leaked=12 MB
After 30 images: leaked=12 MB
After 40 images: leaked=12 MB
After 50 images: leaked=12 MB

Total: compressed=227 MB, leaked=12 MB
Leak per image: 0.2 MB