Optimise NMT Performance Through Buffer Reuse

[mcrakhman]

We can improve the performance of Push and ComputeRoot methods by implementing buffer reuse strategies for hash computations (HashLeaf and HashNode methods).

While this optimisation won't be suitable for all use cases, it can provide performance improvements in ProcessProposal.

Proposed Optimisations:

In-place root computation

Introduce a new method (e.g., RootInPlace, ConsumeRoot, or FastConsumeRoot) that:

• Reuses the leafHashes buffer when computing intermediate node hashes during root calculation
• Modifies internal buffers in place rather than allocating new ones
• Clearly indicates through naming that the NMT instance becomes unusable after the operation

NMT instance pooling

Implement object pooling for ProcessProposal:

• Maintain a pool of pre-allocated NMT instances
• Allocate memory once during the first block processing
• Reuse NMT instances from the pool for subsequent operations, reducing allocation overhead

[tzdybal]

There is definitely a potential to reuse buffers (I feel we should be able to use only 3 buffers for every computeRoot recursive call stack). Currently the code is really easy to understand, reusing those buffers makes code less readable and more complicated, unfortunately.

Regarding pooling, I think this is not an issue for NMT itself - it's more for users of this library.

[tzdybal]

One low-hanging-fruit is to stop allocating and copying data for leaf hashes - we can just return pre-computed values without copying.