Clone zeros in pieces as large as the largest chunk

[folbricht]

When zeros are written over existing data on a filesystem that supports reflinks, the null seed clones them from a temporary file of zeros. That file was one block long, so each 4 KiB block of zeros was a separate FICLONERANGE call and ended up as a separate extent. Blank targets aren't affected since #439, which leaves zeros out of them, but updates over existing data are.

The file of zeros is now as large as the largest chunk, rounded up to full blocks: 256 KiB with the default chunk sizes. Runs of zeros are cloned in pieces of that size, and only the last piece is shorter. The unaligned head and tail are still written as zeros, and a failed clone still falls back to writing zeros.

Measurement

These are Debian 13 cloud images (3 GiB raw, about 2 GiB of it zero chunks) on btrfs, each updated in place to 20260909 with #323 applied (extract -k --seed old.caibx:target). There were two runs per case. Every result matched the image's md5.

Update from		Extents	Wall time
20260907 (2 days)	before	43k – 56k	24.8 – 32.3 s
	after	7.8k – 7.9k	11.2 – 16.5 s
20250909 (1 year)	before	32k – 40k	29.5 – 39.9 s
	after	8.2k	11.5 – 19.8 s

A plain extract over an existing image on master writes only about 65 MiB of zeros, because the rest is already in place. There, the extent count varies between 3k and 10k from run to run with either binary, so that case shows no measurable difference.

The temporary file now takes up ChunkSizeMax bytes instead of one block while an extract runs.