Add non-dup/known-size functionality to the lazy parser for cu optimization

[figaro-sol]

By letting the user inform the lazy account parser about whether the next account is known not to be a duplicate and if the size is known, users can get much better code generation. llvm is rather bad at propagating 'if condition then unreachable_unchecked' backwards. For example, in the CU bench,

let authority = match context.next_account() {
      Ok(MaybeAccount::Account(acc)) => acc,
      Ok(MaybeAccount::Duplicated(_)) => unsafe { core::hint::unreachable_unchecked() },
      Err(_) => hard_exit("Bad authority account"),
};

if authority.data_len() != 0 {
    unsafe { core::hint::unreachable_unchecked() }
}

has at best marginal effect on CUs of the updater (40->39 CUs). The asm generated for account parsing + size/signer includes many dup checks and full offset calculations like

ja +0x21 <entrypoint+0x198>
ldxdw r3, [r1 + 0x58]
add64 r1, 0x8
add64 r2, r3
add64 r2, 0x285f
and64 r2, -0x8
ldxb r3, [r1 + 0x1]
jne r3, 0x0, -0x10 <entrypoint+0x48>

and

mov64 r2, r1
add64 r2, 0x10
ldxb r3, [r1 + 0x8]
jeq r3, 0xff, +0xc <entrypoint+0x90>
mov64 r1, 0x0
ldxb r3, [r1 + 0x1]
jeq r3, 0x0, +0x10 <entrypoint+0xc8>
mov64 r3, r2
add64 r3, 0x8
ldxb r4, [r2 + 0x0]
jeq r4, 0xff, +0x10 <entrypoint+0xe8>

The version using the new gates:

let Ok(authority) = (unsafe {
    context.next_account_guarded(&AssumeNeverDup::new(), &CheckLikeType::<()>::new())
}) else {
    hard_exit("Bad authority account");
};q

used only 26 CUs. The generated code fully constant-propagates the offset computation and removes dup checks giving

ldxdw r2, [r1 + 0x0]
jne r2, 0x2, +0x1c <entrypoint+0xf0>
ldxdw r2, [r1 + 0x58]
jne r2, 0x0, +0x27 <entrypoint+0x158>
ldxb r2, [r1 + 0x9]
jeq r2, 0x0, +0x1c <entrypoint+0x110>
ldxdw r2, [r1 + 0x28b8]
jne r2, 0x28, +0x27 <entrypoint+0x178>
---- authority check, etc

This change also adds assume-stye assertions about the alignment of the instruction data as part of the process. this helps libraries like bytemuck elide alignment checks.

Checked and unchecked variants of the dup/data gates are added as well. the checked variants still provide a significant performance improvement since the compiler can usually infer account offsets with the check in place.

I've included the cu benchmark code I used into the PR. not sure if it's desired or not, happy to remove.