Improve pipelining for validated sending in provider

[rklaehn]

Currently, we use bao_tree::tokio::encode_ranges_validated to encode a blob. This leads to a small but detectable performance decrease, since the operation is not pipelined at all.

It is basically:

• read chunk
• validate chunk
• write chunk

in a loop.

We could pipeline things so that reading happens in a stream, validation is an intermediate stage, and writing is the last stage. Then you can tweak both read parallelism and validate parallelism by just using https://docs.rs/futures/latest/futures/stream/trait.StreamExt.html#method.buffered in the right places.

[rklaehn]

This would still give a small benefit, but once BLAKE3-team/BLAKE3#329 is merged all bao-tree blake3 functions will be even faster, on all architectures that support any kind of SIMD.

So the pressure to do this is quite a bit less...