zstd: make use of RLE_Block

[dsnet]

Consider the following snippet:

bb := new(bytes.Buffer)
zw := must.Get(zstd.NewWriter(bb, zstd.WithEncoderLevel(zstd.SpeedBestCompression)))
must.Get(zw.Write(make([]byte, 10000)))
must.Do(zw.Close())
fmt.Printf("%x\n", bb.Bytes())
// 28b52ffd 64 1026 4d0000 080001540102310c87 0144e310

where we're compressing a input of 1e4 '\x00' bytes.

In the hexdump, we see that a Compressed_Block (type 2) was produced:

4d0000 080001540102310c87

Ideally, the compressor identify this as a candidate for RLE_Block (type 1):

833801 00

which is 8B shorter and faster to parse.

[klauspost]

Correct. There is no RLE detection. It is just such a fringe case it hasn't really come up.

It might be pretty cheap to detect after the compression loop. Let me see if can make it work reliably.

[klauspost]

Added in #938

[klauspost]

@dsnet Output of your example is:

// 28b52ffd 64 1026 4d0000 080001540102310c87 0144e310
-> 
// 28b52ffd 64 1026 83380100 0144e310

[dsnet]

Very cool. Thank you. Sparse files were one of the reasonable use cases that I thought this would be useful for.