Fix stake program panics and failures when testing with mainnet-forked state

[aoikurokawa]

Summary

Fixes three bugs that made the stake program unusable in LiteSVM when
testing against mainnet-forked state (reported in
solana-foundation/surfpool#605).

• Update stake ELF to v5.0.0: The bundled core_bpf_stake-1.0.1.so
  was compiled against solana-vote-interface v2.x and only knew
  VoteStateVersions discriminants 0–2. Delegating to a V4 vote account
  (discriminant 3, the current mainnet format) returned InvalidAccountData.
  Replaced with core_bpf_stake-5.0.0.so.

• Fix StakeHistory account size: StakeHistory::size_of() is
  hard-coded to 16 KiB (512 max entries). Initialising via set_sysvar
  allocated that full padded buffer, so sol_get_sysvar reads beyond
  the actual data returned zeros instead of an error. The Stake BPF
  program asserts entry_epoch == target_epoch after each partial read,
  causing a panic at any epoch >= 1 for operations that read stake
  history (e.g. Merge, Split, DelegateStake). The account is now
  initialised with the exact serialized size so out-of-bounds reads fail
  cleanly.

• Populate StakeConfig account: The deprecated StakeConfig account
  (StakeConfig11111111111111111111111111111111) was never created,
  causing DelegateStake instructions that pass it as an account to
  fail. It is now populated in set_sysvars() with the correct
  config-program encoding (8-byte key-count header +
  bincode(Config::default())).