Bump itoa version, don't use mem::uninit (0.13 backport)

[ghost]

Marking this as a draft for now while I test some crates, patching their hyper version to be this one then seeing if a cargo update fixes them with strict init checks. The code is done, though.

See: #2914

[ghost]

Hm.... I needed to bump pnet for the tests to pass with strict checks (otherwise you run into uninit memory in there), but that means the MSRV of 1.39 fails tests. And looks like a lot of dependencies depend on the 2021 edition in semver compatible versions with the ones we have in cargo.toml

[ghost]

Also, i'm testing according to the checks that are enabled with -Zstrict-init-checks, which means "no uninit integers and no zeroed types that can't be zeroed". This is basically the checks that miri does, but is strictly speaking, stricter than what's been ruled to be UB yet (see rust-lang/rust#98919 ).

I figure it's best to fix all the panics that rust can legally insert now rather than needing to do another backport if/when we get the chance to make let x: [u8; 100] = mem::uninit() panic.

[RalfJung]

This is basically the checks that miri does, but is strictly speaking, stricter than what's been ruled to be UB yet

Well, the reference says it is UB.

[ghost]

Okay, I tested a few crates (arangoq, jqdata, oxidux, usgs-espa-client) that use 0.13 hyper by patching this into their dep tree, and running cargo update if http complained as well, which was only an issue for oxidux.

All 4 compiled and had their tests run fine, even with -Zstrict-init-checks.

[ghost]

Closing this as not really needed anymore, 0.13 is not maintained and there's no risk of anything actually going wrong even in far future LLVM because of the 0x01 filling.