Improve MulEven performance for RVV

[lsrcz]

This pull request improves the performance of MulEven by using masked multiplication instead of merging the results after computing the higher and lower parts.

A similar optimization can be applied to the MulOdd operator, though this would require a MaskedMulHighOr operator, which is currently unavailable.

For 8-bit elements, MulEven currently uses the immediate value 0x5555, which requires two instructions to construct on RISC-V. There may be a potential optimization by constructing the mask manually instead of relying on Dup128MaskFromMaskBits, allowing the use of the smaller immediate 0x55. However, it's just a scalar instruction, so I am unsure if we want to do so.

[jan-wassenberg]

Nice, looks like we're fusing OddEven into the computation. Thanks for sending the PR!
Is this also automatically found?

I would expect scalar instructions are fine, the scalar pipes might be running ahead and/or idle.

[lsrcz]

Yes, a synthesizer automatically finds this. I have no idea why the CI fails though -- it seems that it doesn't come from this pull request.

For MulOdd, do you think adding the masked MulHigh and do the optimization is worthwhile?

[jan-wassenberg]

Nice. I agree CI failures are unrelated and will fix some of that shortly.

I think MulOdd would be rarely used, so let's leave it as-is for now :)