Transmuting `&mut MaybeUninit::zeroed()` to `&mut [u8]` is unsound

## Progress
- [x] Fix unsoundness in `FromBytes::read_from_io`
- #2320
- #2358
- [ ] Update documentation as described below

---

These days, `MaybeUninit::zeroed` is [pretty clear](https://doc.rust-lang.org/std/mem/union.MaybeUninit.html#method.zeroed) that you cannot expect all padding bytes to be zero in the output value (same for [`mem::zeroed`](https://doc.rust-lang.org/std/mem/fn.zeroed.html)). Zerocopy [mentions this](https://docs.rs/zerocopy/latest/zerocopy/trait.FromZeros.html#warning-padding-bytes), but it doesn't seem fully internalized by the library:

- `FromZeros::zero` contradicts this, stating that it's similar to `*self = Self::new_zeroed()` and that the difference is it doesn't semantically drop anything. This isn't the only difference! `zero` is capable of fully zeroing all of the bytes of an object until it is moved into, while `*self = Self::new_zeroed()` doesn't guarantee you can then soundly access as a byte slice unless that `Self: IntoBytes`.
- `FromZeros::new_zeroed` doesn't mention the same caveat as `MaybeUninit::zeroed` at all even though it applies.
- The [current implementation](https://github.com/google/zerocopy/blob/17e7e4d98c9d2cd49c94307e2276b7ce377aa74a/src/lib.rs#L4551) of `FromBytes::read_from_io` is unsound, since there may still be padding bytes for `!IntoBytes` types from `new_zeroed`. It should be constructing a `MaybeUninit::uninit()` then [zeroing its bytes](https://doc.rust-lang.org/std/ptr/fn.write_bytes.html) before casting to `&mut [u8]`.

I'll send a PR fixing the unsoundness issue. I'm also interested in improving the documentation to be more consistent and to provide clear working alternatives to documented unsound operations.