We can get rid of uneeded memsets when clearing state, since
the count is always reset.

We can clear the hasher state on init if we need to start from
zeroes.

Signed-off-by: Pantelis Antoniou <[email protected]>

This is a minimal thread-pool implementation for BLAKE3.

As part of the commit a very simple file mmap-ing example is
introduced that providing file checksuming.

At least on my machines (x86-64 linux boxes) the C implementation
is within a 1% point in performance of the rust b3sum tool.

> $ ls -l ggml-vicuna-7b-1.1-q4_0.bin
-rw-rw-r-- 1 panto panto 3791725184 Ιουν 17 13:34 ggml-vicuna-7b-1.1-q4_0.bin

> $ time b3sum ggml-vicuna-7b-1.1-q4_0.bin
> fe87fecc4427c4661aa46f52a7e286ecd6dd7f5f788564abb2c46d2ed5341584  ggml-vicuna-7b-1.1-q4_0.bin
>
> real	0m0,190s
> user	0m1,583s
> sys	0m0,196s

> $ time ./example-mmap ggml-vicuna-7b-1.1-q4_0.bin
> fe87fecc4427c4661aa46f52a7e286ecd6dd7f5f788564abb2c46d2ed5341584  ggml-vicuna-7b-1.1-q4_0.bin
>
> real	0m0,204s
> user	0m1,525s
> sys	0m0,247s

At my local copy of BLAKE3 with more extensive changes I can get the C
implementation to be slightly faster than the rust one.
But those require more extensive changes which will have to come much
later.

Signed-off-by: Pantelis Antoniou <[email protected]>

When buffer sizes are small in streaming mode the partial output
must have the rest of the buffer filled with zeroes.

Signed-off-by: Pantelis Antoniou <[email protected]>