Build without std, behind a default-on std feature

[pathscale]

Arctic builds without std. The default feature set gains std, so nothing
changes for any existing user: same code paths, same 41 tests, none modified.

What it actually needed std for

Less than the 15,820 lines suggest. Four things, and only one of them was in the
algorithm:

	was	is
the SIMD level	std::sync::LazyLock	spin::Once
a Sync marker	PhantomData<std::sync::Mutex<V>>	a private MutexLike<V> with the two unsafe impls spelled out
the error impl	std::error::Error	core::error::Error, stable since 1.81
everything else	std::fmt, std::cmp, Box, Vec	core:: and alloc::

The marker is the one worth looking at in review. core has no type whose auto
traits are "Send when V: Send, Sync when V: Send", which is what
Mutex<V> gave. Cell and UnsafeCell are the near misses and both are never
Sync, so it is written out with the two impls and a SAFETY note rather than
approximated.

What you lose without std, stated rather than buried

Runtime SIMD detection. is_aarch64_feature_detected! and its x86 twin are
std macros, so fearless_simd::Level::new is gated on std. Without it the
level comes from cfg!(target_feature) at compile time.

On aarch64 that costs nothing: Neon is mandatory in the architecture, so
detection had exactly one possible answer, and it was reaching it through a
Once, an atomic load, and a match at each of the ten dispatch! sites. On
x86 it means the scalar path unless the build asks, with -C target-feature=+avx2
or a target-cpu that implies it. If runtime detection is ever wanted without
std, cpufeatures does it with core::arch::__cpuid and no std.

I checked the compiled artifact rather than trusting the source: 1,049 Neon
instructions, cmeq for the key compares, ushll/ushll2 widening for the
u16x16 sort. The SIMD path is real and present.

Reclamation. smr-ps-reclaim and stat-garbage now imply std, because
ps-reclaim is std only and the garbage counter batches per thread, which is
this crate's only thread_local!. NoOp is what remains available.

Dependencies

fearless_simd needs libm without std, for float maths a radix tree never
calls. That is the sanctioned path rather than a workaround: upstream has an
open draft adding no-std libm tests to their CI (linebender/fearless_simd#229).

Every crate in the no_std graph already declares no_std, so nothing needs
forking:

arctic-wt
├── fearless_simd → libm
└── ribbit        → arbitrary-int, portable-atomic
                  → ribbit-derive (proc-macro, host only)

Verified

cargo check and cargo check --no-default-features both clean, zero warnings
on either. cargo fmt --check clean. 41 tests pass.

Not verified: a genuinely bare target. This repo pins toolchain 1.96.0 in
rust-toolchain.toml and only aarch64-apple-darwin is installed for it, so
cargo check --target thumbv7em-none-eabi fails on a missing core rather than
on anything in this change. One rustup target add closes that, and it is worth
doing before this is relied on.

While in here, one thing that is not in this PR

Scalable vectors are a dead end for this crate, and I would rather write down
why than have it re-derived. This M4 Max reports sme_max_svl_b: 64, a 512-bit
streaming vector length, and node_47's 48-key search is four Neon compares
that would be one. Three separate walls stop it: SVE intrinsics are nightly only
and blocked on an unaccepted RFC for unsized types (rust-lang/rust#145052);
Apple has no non-streaming SVE, and -C target-feature=+sve produces a binary
that dies with SIGILL on this machine, which I measured; and fearless_simd has
refused scalable vectors by design (linebender/fearless_simd#339), because its
whole API rests on a compile-time lane count.