AtomicCell: Use atomic-maybe-uninit

[taiki-e]

This fixes UB when lock-free and the value has uninitialized bytes, by using atomic-maybe-uninit crate.
Fixes #315
Fixes #748

Also, this makes 128-bit atomics lock-free on some platforms.
Closes crossbeam-rs/rfcs#13

Currently, this patch always uses fallback (seqlock) on environments that do not support inline assembly, such as Miri.
It is possible to use atomic-maybe-uninit crate to fix UB of concurrent access with volatile read/write in seqlock/deque (#744, #859), but that is not included in this PR and is left to subsequent PRs.

---

cc @RalfJung

[taiki-e]

UPDATE: resolved in Rust 1.74+ by rust-lang/rust#114790 & taiki-e/atomic-maybe-uninit#19.

Currently, it is not possible to use MaybeUninit in the input registers of an inline assembly (i.e., when processing MaybeUninit, values must be passed through memory), so this worsens performance.

Before:

test concurrent_load_u8     ... bench:     600,658 ns/iter (+/- 11,986)
test concurrent_load_usize  ... bench:     601,385 ns/iter (+/- 11,719)
test concurrent_store_u8     ... bench:   1,358,933 ns/iter (+/- 192,961)
test concurrent_store_usize  ... bench:   1,356,045 ns/iter (+/- 20,230)

After (pass by memory):

test concurrent_load_u8     ... bench:     868,554 ns/iter (+/- 150,633)
test concurrent_load_usize  ... bench:     851,380 ns/iter (+/- 7,139)
test concurrent_store_u8     ... bench:   2,029,710 ns/iter (+/- 178,712)
test concurrent_store_usize  ... bench:   2,038,341 ns/iter (+/- 196,265)

If I change the atomic-maybe-uninit implementation to pass by value using transmute, performance will be equivalent to the current one, but it is UB when the input value contains uninitialized bytes (because it is transmuted to integer).

After (pass by value):

test concurrent_load_u8     ... bench:     601,504 ns/iter (+/- 12,824)
test concurrent_load_usize  ... bench:     600,052 ns/iter (+/- 15,885)
test concurrent_store_u8     ... bench:   1,357,697 ns/iter (+/- 211,465)
test concurrent_store_usize  ... bench:   1,357,504 ns/iter (+/- 252,722)

It would be nice if the inline assembly could allow MaybeUninit registers as thomcc had previously proposed.

[taiki-e]

Filed rust-lang/rust#114790 to allow MaybeUninit in input and output of inline assembly.

[taiki-e]

Since this PR was opened we have made many improvements to both atomic-maybe-uninit and the compiler. Resolving performance issue (rust-lang/rust#114790), improving and stabilizing inline assembly support on tier 2 targets (rust-lang/rust#131258, rust-lang/rust#131781, etc.), fixing or avoiding rustc/LLVM bugs, improving code quality (taiki-e/atomic-maybe-uninit#30, etc.), improving platform support (atomic-maybe-uninit currently supports all CPU architectures supported by Rust other than C-SKY), etc.

I think we can now finally merge this.