Remove bounds checks from color transform hot loop

[Shnatsel]

The code isn't vectorized on baseline x86_64 but at least per-block bounds checks in the inner loop are gone.

Assembly:
before: https://rust.godbolt.org/z/f5r1noP51
after: https://rust.godbolt.org/z/1fx9qebvh

If I tell the compiler to target Zen 3, it does vectorize the code, so perhaps this function would be a good target for multiversioning: https://rust.godbolt.org/z/jrnor5vqY
But that's out of scope of this PR.

[kornelski]

Couple more panics can be removed:

    let block_size = 4 << size_bits;
    if block_size == 0 || width == 0 || block_xsize == 0 {
        return;
    }
    …
    let Some(row_tf_data) = transform_data.get(row_transform_data_start..) else { continue; };

This is probably negligible for code speed, but makes the compiled code nicer.

[Shnatsel]

Another thing that removes a lot of code is turning this

#[inline]
pub(crate) fn subsample_size(size: u16, bits: u8) -> u16 {
    ((u32::from(size) + (1u32 << bits) - 1) >> bits)
        .try_into()
        .unwrap()
}

into this

#[inline]
pub(crate) fn subsample_size(size: u16, bits: u8) -> u16 {
    ((u32::from(size) + (1u32 << bits) - 1) >> bits) as u16
}

which makes the godbolt assembly way shorter.

But I don't think it matters for performance and it's nice to have a sanity check so I left it as is.