spl-token-2022 token-metadata decoder panics on forged Borsh string lengths

[tschifra]

spl-token-2022 accepts token-metadata interface instructions through a dispatch
fall-through to TokenMetadataInstruction::unpack(input) in
program/src/processor.rs:2236-2237. For three of the metadata variants — Initialize,
UpdateField, RemoveKey — the decoder calls try_from_slice(rest) directly on
attacker-controlled instruction-data bytes after matching an 8-byte discriminator
(defined in spl-token-metadata-interface 0.8.0 at src/instruction.rs:170-180). Those
types contain Borsh String fields. A forged u32::MAX length prefix causes Borsh to
attempt a multi-gigabyte allocation against the SBF VM's bounded heap; the program
aborts with ProgramFailedToComplete instead of returning a typed
InvalidInstructionData error.

[bit2swaz]

im taking this one

i traced the current metadata fall-through in program/src/processor.rs and reproduced the malformed-length failure for Initialize, UpdateField, and RemoveKey

ive put together a narrow local fix that pre-validates those recognized metadata payloads before unpack, so malformed inputs now return InvalidInstructionData instead of panicking / ProgramFailedToComplete

im opening a PR shortly