Document oddities in validation behaviour of calldata-to-memory copies.

Arrays with static (as in "not dynamically encoded") base types use an old custom copy routine for calldata-to-memory copies, which does not perform cleanup or validation, while arrays with dynamically encoded base types use the yul conversion routine, which uses abi-decoding and validates, e.g. reverts on dirty higher-order bits (see ``CompilerUtils.cpp:L1033`` for the case distinction).

This fact for legacy code generation should be documented somewhere.

Additionally, via-IR-code generation always implements such copies using abi-decoding, so this difference should be documented as an via-IR-breaking-change.

Came up in https://github.com/ethereum/solidity/issues/13518

Furthermore, once this is documented for now, we should actually change this to *always* validating in all such copies in 0.9.

Start with the case distinction at CompilerUtils.cpp:L1033 and review the related context from issue #13518. Document the legacy code-generation difference between static and dynamically encoded base types, the via-IR validation behavior as a breaking change, and the planned always-validate change for 0.9.

Written by the indexing model from the issue text.