perf: remove unnecessary heap allocs

[malik672]

• Replace push/extend_from_slice loops with pre-sized set_len + offset writes across xmss and rec_aggregation
• Replace .to_vec() / .concat().try_into() with copy_from_slice into stack arrays
• Pre-allocate trace vectors with known capacity

this shows improvement on 1000sig aggregation (120 → 137 sig/s)

pub fn to_big_endian_bits(value: usize, bit_count: usize) -> Vec<bool> {
  (0..bit_count).rev().map(|i| (value >> i) & 1 == 1).collect()
}

pub fn to_big_endian_in_field<F: Field>(value: usize, bit_count: usize) -> Vec<F> {
  (0..bit_count)
      .rev()
      .map(|i| F::from_bool((value >> i) & 1 == 1))
      .collect()
}

can be done here too

[TomWambsgans]

Hi, thanks for your PR. I believe the "120 → 137 sig/s" is likely a statistical fluctuation, since this part of the code should represent a negligible portion of prover's work. That's why I am a bit reluctant to add 50 LoC for an unclear benefit. But I think one can still optimize this without adding all these lines, simply by adding Vec::with_capacityfor instance, what do you think?

[malik672]

Hi, thanks for your PR. I believe the "120 → 137 sig/s" is likely a statistical fluctuation, since this part of the code should represent a negligible portion of prover's work. That's why I am a bit reluctant to add 50 LoC for an unclear benefit. But I think one can still optimize this without adding all these lines, simply by adding Vec::with_capacityfor instance, what do you think?

Thanks
true the prover is the main bottleneck, coming will recheck