feat: DLP v2

[snawaz]

In the first phase of DLP v2, implement the following ix:

• delegate
• commit
• finalize
• undelegate
• call_handler

performance tips

• validations:

  • avoid find_program_address; pass bumps.
  • avoid create_program_address if possible; consider using derive_address instead.
  • minimize key comparisons.
  • fail early, cheap checks first.

• memory & data:

  • avoid heap allocation; use zero-copy aggressively.
  • avoid copies as well for bigger types (> 8 bytes).
  • avoid wide arithmetic (u128 / U256) in hot path.
  • prefer fixed-size account layouts.
  • smaller binary: try to reuse code.

• accounts:

  • merge multiple accounts where possible.
  • minimize total loaded account data size.
  • avoid creating / reallocating accounts in commit/finalize.
  • move cold data out of hot-path accounts.

• PDAs & signers:

  • avoid PDA signing unless strictly necessary.
  • cache derived PDAs in state when reused.

• instructions:

  • minimize instruction count, especially on commit path.
    • avoid CPI in commit/finalize.
    • avoid ComputeBudget reliance.
    • prefer one large ix instead of many smaller ix.
  • commit/finalize must be infinitely boringly simple, even if setup is complex.
    • single instruction only.
    • no helper instructions.
    • no multi-step commit flow.
    • no retry logic inside the program.

• logging:

  • zero logs on success
  • detailed logs on failure (including ix discriminator)

• build strategy:

  • feature-gate debug / validation logic
  • avoid abstractions that prevent inlining in hot path