New program: serialization library study

[joncinque]

Problem

There's mostly anecdotal data about the tradeoffs of different serialization techniques used by on-chain programs. The typically used ones are:

• Pack by hand, as in token
• unsafe pointer casts
• Borsh, as in governance and stake-pool, among others
• bytemuck, as in certain token-2022 extensions
• bincode, not used in BPF programs, but builtin programs in the monorepo use it, such as stake
• protobuf, a bit more overhead
• any others?

Solution

Add in a new libraries/serialization style program that compares these different approaches. #3064 was the most bare-bones prototype possible that can be used for inspiration. Some features to compare them on would be:

• compute unit usage for different structs / enum (for example, bincode is known to have high compute usage)
• bloat of a compiled program (for example, borsh is known to make a compiled program huge)
• ease of use (for example, bytemuck only allows for types that can be represented as bytes)

[alessandrod]

Related issue: solana-labs/solana#23075. solana_prorgam unconditionally pulls in both borsh and bincode, so until that's fixed, all on chain programs that depend on solana_program will be somewhat bloated.