Add safe version of read_u16_le() that is exactly as fast

[Shnatsel]

Add safe version of read_u16_le() that is exactly as fast as unsafe implementations, at least on x86_64.

Sadly, it only works on Rust 1.34 and later due to use of TryFrom trait, so it is commented out for now.

I am not aware of any conditional compilation support based on rustc version, other than https://github.com/dtolnay/select-rustc which pulls in syn as a dependency which considerably increases compilation time in release mode.

[Shnatsel]

https://github.com/cuviper/autocfg is be a syn-less way to check for compiler features (it uses build.rs) but I have almost zero experience with build.rs so I will not integrate it myself

[oyvindln]

I don't know whether 1.32 is old enough to be fine to require. But if so, we should replace all these.

Did you also check the performance of doing it via a bit-shift like in the big-endian version? Maybe the compiler would be smart enough to optimize that now.

[Shnatsel]

This actually requires 1.34 because of TryInto. I don't think it's old enough to require yet, but will come to Debian Stable this year, at which point I think it will be fine to require it.

I did not check the big-endian version since I do not have any big-endian hardware. But I'd expect u16::from_le_bytes to perform well on big-endian, or at least as good as anything we can hand-roll.

[oyvindln]

I meant to see if implementing it with bit shifts rather than reading a [u8;2] as a u16 would be as fast on little endian. On big endian we already do it with bit shifts so I expect u16::from_le_bytes to be at least as fast if not faster.

[oyvindln]

Debian stable just had a new release, and now has rust 1.34, so it would probably be fine to adopt the new features and avoid some more unsafe code.

[Shnatsel]

Ubuntu LTS is still stuck on 1.32 though, and this PR requires 1.34

[oyvindln]

Closing this as I removed the unsafe usage in this function in the last commit without needing try_into.