[stateless validation]: Implement state witness distribution by parts

[saketh-are]

Currently the chunk producer has to send the state witness V times, where V is the number of chunk validators. Delivery of the state witness is latency sensitive as it is necessary for block production.

In a case with a sizable state witness (say, 8 MB) and 50 chunk validators, it means that the chunk producer needs to urgently send a burst of 400 MB, which overwhelms the local capabilities of the node.

By implementing a distribution mechanism based on splitting the witness into parts and relying on other validators to forward parts, we can spread the work among all of the chunk validators. It is similar to what is done in mainnet today for chunk distribution.

We plan to:

• Use Reed-Solomon encoding to break the state witness into V parts, where V is the number of chunk validators
• For each part, assign one chunk validator to be responsible for forwarding the part to all other chunk validators
• Decode the state witness when enough parts have arrived and proceed as normal

We expect the latency of this 2-hop approach to be acceptable because it works well in production today for chunk distribution via PartialEncodedChunks.

After this change, each node will only need to send an amount of data equal to the size of the state witness, plus a 50% overhead from Reed-Solomon encoding. As an added benefit of the erasure coding, each validator only needs to receive 2/3 of the parts before the state witness can be decoded.

[shreyan-gupta]

Some PRs for the task https://github.com/near/nearcore/pull/11079 https://github.com/near/nearcore/pull/11080 https://github.com/near/nearcore/pull/11089 https://github.com/near/nearcore/pull/11090 https://github.com/near/nearcore/pull/11091

[shreyan-gupta]

https://github.com/near/nearcore/pull/11115

[walnut-the-cat]

are we done with this work?