Dilithium on-the-fly matrix generation for keygen and verify

[dop-amin]

Reduces stack usage in keygen and verify while no impact on the cycle count is to be expected.

[mkannwischer]

Looks good to me.
As expected there is no significant change in cycles, but key gen and verification stack usage are a lot smaller.
Thanks, @dop-amin!

[rpls]

Should we merge this as a separate implementation? Something like m4f-memopt?

[dop-amin]

Superseded by #340.