Support removing blocks

[Philipp Hanslovsky]

Currently, the N5Writer interface does not expose an interface to delete a block. Tools that copy or convert data with the option to overwrite have to write all blocks of a data set, even if they are "empty" (e.g. all zero), e.g. saalfeldlab/n5-spark#12. Not writing out "empty" blocks is in particular useful for large datasets with substantial empty areas as brought up in saalfeldlab/paintera-conversion-helper#33. I propose to add support for removing blocks at the N5Writer level:

N5Writer.deleteBlock(String, DatasetAttributes, long... blockPosition);

Another option would be to implicitly delete blocks if a special implementation of the DataBlock is passed to indicate deletion. Unfortunately, we cannot pass null for the DataBlock because the block also holds the blockPosition.

I personally prefer the explicit option (minor version bump) but I would like to hear opinions before I start working on this.

cc @aschampion @igorpisarev

[Igor Pisarev]

I like your suggestion about adding a separate method for it, especially since there is no easy way to integrate it into writeBlock() as you pointed out.

[Philipp Hanslovsky]

This is becoming increasingly more important, overwriting existing files is slower than writing new files, e.g. on the SSD in my workstation:

$ dd if=/dev/zero of=test bs=4k count=$((128*2000))
256000+0 records in
256000+0 records out
1048576000 bytes (1.0 GB, 1000 MiB) copied, 1.17341 s, 894 MB/s
$ dd if=/dev/zero of=test bs=4k count=$((128*2000))
256000+0 records in
256000+0 records out
1048576000 bytes (1.0 GB, 1000 MiB) copied, 58.83 s, 17.8 MB/s
$ rm test
$ dd if=/dev/zero of=test bs=4k count=$((128*2000))
256000+0 records in
256000+0 records out
1048576000 bytes (1.0 GB, 1000 MiB) copied, 1.16586 s, 899 MB/s

This is consistent every time I try and repeat and probably affects N5 writes as well (albeit not as big). If we can delete blocks, we can alwayas remove blocks before writing.

Update: This is relevant for ext4