Fix InstructionHolder scratch ownership

[edo-chan]

Summary

This PR is stacked on codex/sign-v2-cu-comparison and fixes the InstructionHolder scratch lifetime issue without adding heap allocation.

Key changes:

• replace returned slices into parser-local MaybeUninit arrays with a lending parser API: InstructionIterator::process_next
• add InstructionScratch fixed buffers for AccountMeta, original account indexes, and CPI Account values
• execute/inspect each parsed instruction inside the callback while scratch-backed slices are still valid
• update SignV2, SOL/token destination parsing, and SubAccountSignV1 call sites to use the fixed-scratch callback flow
• keep parse errors mapped at call sites while preserving callback permission/execution errors
• add a focused InstructionHolder borrow test for scratch-backed account metadata

CU Benchmark

cargo test -p swig --test sign_v2_cu_comparison -- --nocapture

Case	CU
All	2856
SOL limit only	4125
SOL limit + destination	4760
Token limit only	9559
Token limit + destination	10386

These match the current codex/sign-v2-cu-comparison results, so the fixed-scratch rewrite did not increase the LiteSVM CU harness numbers.

Validation

• cargo fmt
• cargo test -p swig-compact-instructions
• cargo test -p swig --test sub_account_test test_sub_account_sign -- --nocapture
• cargo test -p swig --test sign_v2 test_sign_v2_transfer_sol -- --nocapture
• cargo test -p swig --test sol_destination_limit_v2 --test token_destination_limit_v2 -- --nocapture
• cargo test -p swig --test sign_v2_cu_comparison -- --nocapture
• git diff --check
• cargo build-sbf --arch v1

[edo-chan]

Rebased this on top of codex/sign-v2-cu-comparison and replaced the old Vec<AccountMeta> / Vec<usize> approach with fixed scratch for all parsed instruction buffers, including cpi_accounts.

CU rerun on the updated branch:

Case	CU
All	2856
SOL limit only	4125
SOL limit + destination	4760
Token limit only	9559
Token limit + destination	10386

These match the current codex/sign-v2-cu-comparison numbers, so the fixed-scratch rewrite did not increase the LiteSVM CU harness results.

Validation run:

• cargo fmt
• cargo test -p swig-compact-instructions
• cargo test -p swig --test sub_account_test test_sub_account_sign -- --nocapture
• cargo test -p swig --test sign_v2 test_sign_v2_transfer_sol -- --nocapture
• cargo test -p swig --test sol_destination_limit_v2 --test token_destination_limit_v2 -- --nocapture
• cargo test -p swig --test sign_v2_cu_comparison -- --nocapture
• git diff --check
• cargo build-sbf --arch v1

[tracy-codes]

@edo-chan since this is a code change to the instruction parser, let's also test this branch with the litesvm test suite of the typescript sdk to ensure the typescript sdk remains compatible with the program and there's no edges or regressions.