Optimize Multi-threaded implementation of `encode` and `recode`

[itzmeanjan]

In current multi-threaded implementation (i.e. kicks in when parallel feature is enabled) of Encoder::code_with_coding_vector function, we are using rayon parallel iterators to go over each piece. But it is slower than single-threaded SIMD implementation 😮‍💨.

We can instead just run #-of-logical cores many rayon threads and distribute all the pieces among them. It should do better.

rlnc/src/full/encoder.rs

Lines 175 to 222 in 061ec3f

	coded_data.copy_from_slice(
	&self
	.data
	.par_chunks_exact(self.piece_byte_len)
	.zip(coding_vector)
	.map(|(piece, &random_symbol)| {
	#[cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))]
	{
	let mut scalar_x_piece = piece.to_vec();
	gf256_inplace_mul_vec_by_scalar(&mut scalar_x_piece, random_symbol);
	scalar_x_piece
	}
	#[cfg(not(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64")))]
	{
	piece.iter().map(move |&symbol| (Gf256::new(symbol) * Gf256::new(random_symbol)).get())
	}
	})
	.fold(
	|| vec![0u8; self.piece_byte_len],
	|mut acc, cur| {
	#[cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))]
	gf256_inplace_add_vectors(&mut acc, &cur);
	#[cfg(not(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64")))]
	acc.iter_mut().zip(cur).for_each(|(a, b)| {
	*a ^= b;
	});
	acc
	},
	)
	.reduce(
	|| vec![0u8; self.piece_byte_len],
	|mut acc, cur| {
	#[cfg(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64"))]
	gf256_inplace_add_vectors(&mut acc, &cur);
	#[cfg(not(any(target_arch = "x86", target_arch = "x86_64", target_arch = "aarch64")))]
	acc.iter_mut().zip(cur).for_each(|(a, b)| {
	*a ^= b;
	});
	acc
	},
	),
	);

But is it actually? Explore that.