Optimize compression input copy

[malik672]

Summary

• replace two slice copies with a single flattened contiguous copy in symetric::compress
• keep the change scoped to crates/backend/symetric/src/compression.rs

IR / asm impact

For the concrete optimized instantiation compress::<KoalaBear, Poseidon1KoalaBear16, 8, 16>:

Old input-copy shape:

• LLVM IR showed two separate input-side copies: memcpy 32 + memcpy 32
• AArch64 asm loaded/stored the two halves separately before the compression call

New input-copy shape:

• LLVM IR shows one contiguous input-side copy: memcpy 64
• AArch64 asm still lowers that 64-byte copy into two 32-byte vector load/store pairs on this target, but the extra staging buffer in the old helper goes away

Overall optimized function effect:

• old IR: two 32-byte copies into state, then an extra 64-byte copy into the compression buffer
• new IR: one 64-byte copy directly into the compression buffer
• old asm frame size: 224 bytes
• new asm frame size: 160 bytes

Source change:

state[..2 * CHUNK].copy_from_slice(input.as_flattened());

Validation

• cargo check -p mt-symetric