feat(simd): add runtime CPU feature detection for x86_64 using multiversion

[utkarshgupta137]

What type of PR is this?

feat: A new feature

Check the PR title.

• This PR title match the format: <type>(optional scope): <description>
• The description of this PR title is user-oriented and clear enough for others to understand.
• Attach the PR updating the user documentation if the current PR requires user awareness at the usage level.

(Optional) More detailed description for this PR(en: English/zh: Chinese).

Previously, all SIMD feature selection was done at compile time via #[cfg(target_feature = "...")]. This meant binaries compiled without -C target-cpu=native (e.g., for distribution) fell back to scalar code on x86_64, leaving significant performance on the table even on machines with AVX2/PCLMULQDQ support.

This PR adds runtime CPU feature detection for x86_64 using the multiversion crate. The changes are:

Runtime dispatch for get_nonspace_bits and prefix_xor (src/util/arch/):

• When compiled with -C target-cpu=native or explicit feature flags: zero overhead — the runtime detection module is not compiled at all, and the optimized implementations are used directly (identical to before).
• When compiled without feature flags on x86_64: multiversion detects AVX2/PCLMULQDQ support at runtime (once, cached in a static atomic) and dispatches to either the optimized x86_64 implementations or the scalar fallback.

Fix sonic-number SSE2 gating (sonic-number/src/arch/mod.rs):

• The x86_64 SIMD number parsing (simd_str2int) only uses SSE2 intrinsics but was incorrectly gated behind target_feature = "avx2". Relaxed to target_feature = "sse2", which is baseline on all x86_64 CPUs. This means SIMD-accelerated number parsing is now always available on x86_64 regardless of compiler flags.

Performance impact (binaries compiled without -C target-cpu=native, running on AVX2 CPU):

Function	Before	After
get_nonspace_bits	Scalar byte-by-byte loop	AVX2 shuffle (runtime detected)
prefix_xor	Scalar shift cascade	PCLMULQDQ carryless multiply (runtime detected)
simd_str2int	Scalar loop	SSE2 SIMD (always on x86_64)

The sonic-simd abstract types (u8x32, u8x64, StringBlock, etc.) continue to use compile-time selection. When AVX2 is not a compile-time feature, u8x32 is emulated as 2×SSE2 u8x16 — still much faster than scalar. I intend to raise a followup PR for that.

(Optional) Which issue(s) this PR fixes:

Addresses the "runtime CPU detection" item from ROADMAP.md.

(optional) The PR that updates user documentation:

N/A — no user-facing API changes. The runtime detection is transparent and automatic.